ShardLoom Python CLI Client

This package is the first thin Python surface for ShardLoom, a Vortex-native, no-fallback, evidence-certified local compute engine. It invokes the workspace shardloom CLI with --format json, parses the stable OutputEnvelope, and preserves typed result/artifact/certificate payloads, diagnostics, fallback status, and the temporary legacy field mirror.

It is intentionally not a native binding, broad DataFrame API, broad SQL runtime, UDF runtime, or fallback execution path. Importing the package has no ShardLoom side effects. Work happens only when a caller explicitly invokes a CLI command through ShardLoomClient or one of the scoped Python helpers that wraps an evidence-backed CLI smoke.

Public status is owned by docs/release/public-status-matrix.md. This README may describe scoped local Python surfaces, the current source version, and the approved package track, but it does not authorize production support, performance claims, Spark displacement, or hidden external execution.

Local Use

From the repository root:

$env:PYTHONPATH = "python\src"
python -c "from shardloom import ShardLoomClient; print(ShardLoomClient.from_repo().status().status)"

Or install the source-tree package in editable mode for notebook, job, or Foundry-style imports:

python -m pip install -e python

The source package exposes the current technical-preview source version through shardloom.__version__. Public package channels can lag the source tree until the release contract is advanced, so source checkouts should use editable installs and published-channel users should install the latest released package from the selected channel. These channels are install access only and do not imply production readiness, broad runtime support, or performance claims.

python -m pip install shardloom

Published supported-platform wheels resolve the packaged CLI before falling back to PATH. Explicit binary/env/source configuration still wins. Use SHARDLOOM_BIN only when you want to pin a specific CLI binary or when installing from a source distribution without a bundled platform CLI:

$env:SHARDLOOM_BIN = "target\release\shardloom.exe"

Or pass an explicit binary:

from shardloom import ShardLoomClient

client = ShardLoomClient(binary="target/release/shardloom")
print(client.status().status)

ShardLoomClient.from_repo() looks for target/release/shardloom and then target/debug/shardloom when a command is invoked. It does not run commands or probe the repository at import time.

ShardLoomClient.from_env() is the import-friendly constructor for managed Python environments. It reads configuration only and does not run commands:

from shardloom import ShardLoomClient

client = ShardLoomClient.from_env()
smoke = client.smoke_check()
print(smoke.commands)
print(smoke.deployment_capabilities.field("surface_components"))
print(smoke.fallback_attempted)

Supported environment variables:

• SHARDLOOM_BIN: explicit shardloom CLI binary path.
• SHARDLOOM_REPO_ROOT: source checkout containing target/<profile>/shardloom.
• SHARDLOOM_PROFILE_ORDER: comma-separated target profile order, for example release,debug.
• SHARDLOOM_TIMEOUT_SECONDS: per-command subprocess timeout.

If no CLI binary is available, explicit client commands raise ShardLoomBinaryNotFoundError with installation/configuration guidance instead of leaking a raw subprocess error. The exception carries deterministic no-fallback diagnostics plus a shardloom.output.v2-shaped error payload via to_error_payload(command) for agents and wrappers that need protocol-shaped missing-binary evidence. Importing the package and constructing ShardLoomClient.from_env() remain side-effect-free.

For the CG-21 user workflow surface, use shardloom.context() when you want a short import-friendly entry point for smoke checks and capability discovery:

import shardloom as sl

ctx = sl.context()
smoke = ctx.smoke_check()
capabilities = ctx.capabilities()

print(smoke.python_package_version)
print(smoke.resolved_cli_path)
print(smoke.protocol_version)
print(smoke.fallback_attempted)
print(capabilities.python.field("scope"))
print(capabilities.sql_support.capability_state)
print(capabilities.fallback_attempted)

Constructing the context does not run ShardLoom, inspect datasets, probe object stores, touch catalogs, execute SQL, or invoke external engines. The explicit smoke_check() and capabilities() methods run only no-dataset CLI JSON commands and preserve no-fallback status.

Capability views also expose a normalized posture object so Python callers can inspect support, claim, runtime, effect, and policy state without scraping raw CLI text:

posture = capabilities.sql_support.posture

print(posture.support_status)
print(posture.claim_gate_status)
print(posture.report_only, posture.unsupported, posture.claim_grade)
print(posture.runtime_execution)
print(posture.data_read, posture.write_io, posture.object_store_io)
print(posture.fallback_attempted, posture.external_engine_invoked)
print(posture.required_evidence)

The posture view does not widen runtime support. It is a typed convenience surface over existing OutputEnvelope fields and diagnostics. Unsupported or report-only scopes remain unsupported or report-only, and fallback_attempted=false / external_engine_invoked=false stay visible.

Use ctx.user_surface_graduation_matrix() to decide whether a Python or CLI surface belongs on the ergonomic context path. The matrix uses five postures: high_level_context, client_only, diagnostic_only, feature_gated, and not_user_facing. high_level_context rows are the scoped workflows promoted for normal context use; client_only rows stay explicit lower-level CLI/client access; diagnostic_only and feature_gated rows must not be described as runtime support without the matching evidence.

For normal Python use, start from the simple context and query surface. repo_root and profile_order are optional development configuration overrides, not arguments users should have to put in ordinary application code. Source-tree or CI runs can set SHARDLOOM_BIN or SHARDLOOM_REPO_ROOT in the environment when the CLI is not on PATH.

ctx.read(path) is the normal public read wrapper. It infers .csv, .json, .jsonl, .ndjson, .parquet, .arrow, .ipc, .feather, .avro, .orc, and .vortex local source adapters from the path extension. Explicit helpers such as read_csv(...), read_json(...), read_parquet(...), read_arrow_ipc(...), read_avro(...), and read_orc(...) remain available for compatibility, tests, and schema-pinned examples. The context returns a lazy query that uses the shared native workflow for admitted work. Collection returns the complete typed result; a write executes the same query into the declared sink:

import shardloom as sl

ctx = sl.context()
result = (
    ctx.read("target/orders.csv")
    .filter(sl.col("amount") >= 10)
    .select("id", "amount")
    .limit(100)
    .collect()
)

print(result.status)
print(result.result_rows)
print(result.fallback_attempted, result.external_engine_invoked)

VortexWorkflowExecutionReport exposes the result envelope and, for writes, fields such as output_path, rows_written, output_commit_status, and native_io_certificate_status. Use status, fallback_attempted, external_engine_invoked, and claim_gate_status for execution posture. Exact fields depend on the operation and sink; report-only capability views are not runtime evidence.

The same query shape can read admitted local formats through ctx.read(...) or the explicit format helpers. CSV, flat JSON/JSONL/NDJSON, generated rows, and scoped local Vortex inputs are the default public examples. Parquet, Arrow IPC/Feather, Avro, and ORC are admitted scoped local-format surfaces when the matching feature-gated build is present; builds without those readers return deterministic adapter blockers instead of invoking another engine. Compatibility exports such as write_json(...), write_jsonl(...), write_csv(...), feature-gated Parquet/Arrow IPC/Avro/ORC writers, write_vortex(...), and fanout use the same public workflow. Admission depends on the selected input, operation, output format, and enabled build features; unsupported combinations return a deterministic report. Format-specific behavior belongs at read/ingest and write/sink boundaries only; compute semantics should lower through the shared ShardLoom/Vortex runtime or return a deterministic unsupported report. Agents and automation should use docs/reference/shardloom-user-surface-index.md and docs/reference/shardloom-user-surface-index.json as the canonical map of Python, SQL, CLI, generated-source, materialization, and deterministic blocker surfaces. The canonical local output/sink scope is docs/architecture/v1-local-output-sink-scope.md; inspect it with ctx.local_output_sink_scope_report() before treating a write helper as broader than its scoped local evidence.

SQL and DataFrame collection, run(), route() and local writers accept the same memory_gb and max_parallelism request. Writer aliases preserve these settings. Flat aggregate chains retain filters, grouping, measures, HAVING, ordering and limits through the same native admission for collection and all eight writers. Declared compatibility schemas are preserved, including files whose names do not identify their format. SQL NULLS FIRST/NULLS LAST and DataFrame sort(..., nulls="first") or nulls="last" place nulls independently of ASC/DESC. Small aggregate collection returns every row within 65,536 rows and 8 MiB of JSONL, including escaping; larger complete results use the existing writers. For admitted ordering, spill explicitly permits temporary native Vortex runs in an existing absolute local directory:

spill = {"workspace": "/tmp/shardloom-query-work", "quota_bytes": 64 << 20,
         "buffer_bytes": 2 << 20}
ordered = ctx.read_vortex("shipments.vortex").limit(250_000).sort(
    "priority", "label", nulls="last",
)
ordered.write_parquet("ordered.parquet", memory_gb=1, max_parallelism=2, spill=spill)

Create the workspace before execution. route() validates the declaration without probing or creating it. Composed relational order uses buffer_bytes as a retained input flush threshold within one query memory grant; specialized numeric sort and aggregate providers use their existing operator-memory admission. Native sort spill also applies to the flat typed keys scoped by the typed key contract. Supported keys, minimum buffers and spill families remain provider-specific. A spill request does not enable other relational state spill or fanout, relax collection limits, or establish an RSS bound. Successful writes require verified spill cleanup before publication. See docs/reference/native-query-spill.md for exact contracts.

Bounded materialization is explicit. Local-source workflows can carry a limit(...) or pass collect(limit=...); SQL workflows can also pass collect(limit=...) or chain .limit(...).collect(). Those admitted routes return typed report rows from the ShardLoom CLI envelope. Decoded Python-object, pandas, Arrow, NumPy, and notebook materialization helpers are bounded container/output boundaries over the admitted ShardLoom result; optional packages are never used as execution engines and missing packages return deterministic diagnostics:

preview_report = (
    ctx.read("target/orders.csv")
    .select("id", "amount")
    .limit(20)
    .collect()
)
print(preview_report.result_rows)

rows = ctx.read("target/orders.csv").select("id").limit(20).to_python_objects()
print(rows)

pandas_view = ctx.read("target/orders.csv").select("id").limit(20).to_pandas(check=True)
print(pandas_view)

For workflows that need caller-scoped reuse evidence, ctx.session(...) and sl.session(...) expose the same local read/SQL shapes as session-bound workflows:

with ctx.session(session_id="orders-run") as sess:
    result = (
        sess.read_csv("target/orders.csv")
        .select("id", "amount")
        .limit(100)
        .collect()
    )
    repeat = sess.sql("SELECT id FROM 'target/orders.csv' LIMIT 100").collect()
    print(result.reuse_hit, repeat.source_state_reuse_hit)

The session is explicit, caller-owned, and closeable. Its client owns a native worker that can reuse admitted source and prepared query state. Every collection executes the query again; query answers are not cached. Collection, writes and fanout use the same native workflow. Explicit ctx.prepare_vortex(...) calls also track source and prepared-artifact fingerprints for reuse. Session lifetime and reuse evidence do not establish a performance claim.

Supply memory_gb and max_parallelism on each operation that needs an explicit allocation. Session collection, counts, all write_* methods and fanout forward those values to the same native runtime as standalone workflows. Session reuse requires the resource request to match, so changing the allocation cannot reuse an earlier operation's result report. Positive environment defaults SHARDLOOM_MEMORY_GB and SHARDLOOM_MAX_PARALLELISM are read at Python import; an explicit CPU value of 1 stays 1. The built-in defaults remain 4 GiB and 2.

The runtime selects CPU concurrency within the supplied maximum and the CPU capacity available to the process. Ingestion shares that grant among ready source, conversion and writer work, with memory-admitted task windows. This works for arbitrary positive allocations; P4/P6/P8 are examples. I/O, serial readers and memory constraints can limit useful concurrency. The accounted memory budget does not include every upstream allocation or establish a process RSS limit.

For the CLI-visible session lifecycle proof, ShardLoomClient.session_cache_smoke() runs session-cache-smoke --format json and returns a typed SessionCacheSmokeReport. That smoke exercises scoped SourceState, VortexPreparedState, OutputPlan, schema-cache, dictionary-cache, fingerprint invalidation, scratch-buffer reuse accounting, optimizer-trace linkage, explicit close, and cleanup evidence. It is local and claim-gated; persistent cross-process cache, object-store/table reuse, and non-local workflow reuse remain outside this scoped session surface.

The explicit prepare-once Vortex lifecycle is available for advanced validation through a feature-gated CLI/Python surface. Build the CLI with --features vortex-write, then call ctx.read_csv(...).prepare_vortex(workspace=...) for a prepared source, ctx.read_csv(...).prepare_vortex(workspace=...).query(...).collect() for the public Prepare-Once First Query route, ctx.from_rows(...).prepare_vortex(workspace=...), ShardLoomClient.vortex_prepare(...), or ctx.prepare_vortex(...) when you intentionally need to inspect the UniversalIngress -> SourceState -> vortex_ingest -> VortexPreparedState boundary:

@"
id,label,amount
1,alpha,8
2,beta,15
"@ | Set-Content -Encoding utf8 target\vortex-ingest-source.csv

cargo run -q -p shardloom-cli --features vortex-write -- `
  vortex-prepare target\vortex-ingest-source.csv target\vortex-ingest-source.vortex `
  --allow-overwrite --format json

$env:PYTHONPATH = "python\src"
$env:SHARDLOOM_REPO_ROOT = "."
python -c "from shardloom import context; ctx=context(); r=ctx.read_csv('target/vortex-ingest-source.csv').prepare_vortex(workspace='target/shardloom-prepared', allow_overwrite=True); print(r.vortex_ingest_status, r.prepared_state_created, r.prepared_state_reuse_hit, r.prepared_state_reuse_reason, r.fallback_attempted, r.external_engine_invoked)"

Default CLI builds return a deterministic feature-gate blocker instead of writing an artifact. This path is a local fixture smoke; it is not the primary user API, broad Vortex writer support, object-store/table output support, production SQL/DataFrame support, or a performance claim. LazyFrame.prepare_vortex(...) is the higher-level local auto source front door: it derives <workspace>/<source-stem>.vortex when a workspace is supplied, calls the real Rust vortex-prepare route, and exposes prepared_state_reuse_hit, prepared_state_reuse_reason, prepared_state_reuse_manifest_digest, and prepared_state_invalidation_reason through typed properties. It prepares the raw local source before query operators; use .write_vortex(...) when the desired artifact is a query-result sink. Generated-source prepare_vortex(...) uses the existing generated-source Vortex writer and returns a GeneratedSourceWriteReport with prepared_state_created and manifest-backed reuse fields. Repeated compatible generated-source preparation reuses the caller-owned local .vortex artifact through the artifact-adjacent manifest, reports prepared_state_reuse_hit=true, and skips the writer/reopen path when schema, row payload, plan, policy, and artifact fingerprints still match. The route capability report exposes both public prepared front doors as machine-readable rows:

routes = ctx.user_route_capability_report()

for row in routes.public_front_door_route_rows:
    print(row.front_door_id, row.public_user_surface, row.prepared_state_reuse_scope)

Those rows are route guidance and release-readiness evidence. They do not run a benchmark or allow performance, production, or Spark-replacement claims. The benchmark publication bundle mirrors them as public_front_door_benchmark_rows, where they are route-identity rows rather than timing rows. The website uses those rows to show each public Python prepared front door beside its owning route lane, timing boundary, reuse manifest scope, and no-fallback evidence. When capillary preparation is admitted, the report exposes vortex_capillary_preparation_prewrite_status, vortex_capillary_preparation_prewrite_scheduler_applied, and pre-write gate fields for array build, write, reopen, and sink evidence so Python callers can see whether PulseWeave-shaped work windows affected the local route before artifact creation.

For one concrete request, use the public workflow facade. route() is side-effect-free: it does not read the input, write outputs, run SQL, or invoke external engines. run() and prepare(...) execute only admitted ShardLoom-native wrapper paths and attach the same route metadata to the runtime or preparation envelope:

sql_route = ctx.sql("SELECT id FROM 'target/orders.csv' LIMIT 10").route()
df_route = ctx.read("target/orders.csv").select("id").limit(10).route()
execution = ctx.read("target/orders.csv").select("id").limit(10).run()
prepared = ctx.read_csv("target/orders.csv").prepare("target/orders.vortex")
native_vortex = ctx.client.public_workflow_run(
    "cli",
    input_uri="shardloom-vortex/tests/fixtures/local_primitive_struct_five.vortex",
    input_format="vortex",
    requested_output="collect",
    execution_policy="native_vortex",
    materialization_policy="zero_decode",
    evidence_level="runtime_smoke",
    bounded=True,
    vortex_primitive="filter_project",
    vortex_predicate="gte:value:3",
    vortex_columns=("metric",),
    vortex_source_order_limit=2,
    memory_gb=1,
    max_parallelism=2,
)

print(sql_route.route_id, sql_route.resolved_internal_command)
print(df_route.route_id, df_route.resolved_internal_command)
print(sql_route.fallback_attempted, sql_route.external_engine_invoked)
print(sql_route.side_effect_free, sql_route.blocker_id)
print(execution.facade_command, execution.route_id, execution.runtime_execution)
print(prepared.facade_command, prepared.route_id, prepared.preparation_included)
print(native_vortex.command, native_vortex.route_id, native_vortex.vortex_primitive)

For direct .vortex inputs, route() and run() infer the admitted primitive/provider payloads for scoped count/filter/project/limit, no-argument row-level distinct, bounded source-order tail, deterministic row-count sample(n=..., seed=...|random_state=<int>, weights="<numeric-column>", replace=False|True) or fractional sample(frac|fraction=..., seed=...|random_state=<int>, weights="<numeric-column>", replace=False|True), and exact benchmark-family grouped aggregate, hash join, global top-N, cast/try-cast, substring contains, and native write_vortex sink shapes. Manual vortex_primitive and native_vortex_provider_scenario arguments remain available on ctx.client.public_workflow_* for low-level diagnostics, but normal Python/SQL facades route the admitted shapes without requiring those flags.

Unbounded collect requests block at route admission and keep runtime_execution=false, fallback_attempted=false, and external_engine_invoked=false in the envelope. The equivalent CLI surfaces are shardloom route <sql|python|dataframe|cli> --format json, shardloom run <sql|python|dataframe|cli> --format json, and shardloom prepare <sql|python|dataframe|cli> --format json. Lazy DataFrame bounded collect() and admitted generated-source/source-free writes route through the public run facade and return existing typed reports with attached public_workflow_* route fields. write_vortex(...) remains the highest-fidelity native sink when the upstream provider route is admitted. Exact provider-backed result summaries can export bounded result_json to workspace-safe JSONL/CSV. Ordinary SQL/generated local result sinks and primitive filter/project/filter-project row streams also admit JSON arrays; scoped primitive filter/project/filter-project/distinct/tail/sample row streams can export JSONL/CSV or JSONL+CSV fanout through native_vortex_primitive_row_export. JSON and JSONL text outputs do not preserve static type or Vortex layout metadata. Broader compatibility writes such as arbitrary write(...), structured write aliases, unsupported formats, and unsafe or non-admitted fanout return deterministic blockers until a native Vortex sink/export route exists for the normalized plan. Native Vortex primitive and promoted provider helpers attach the inferred route payloads to the same facade rather than relying on a separate payload-only path.

Admitted native DISTINCT, drop_duplicates, duplicated, tail, sample, scalar rewrites, melt, explode, rolling and pivot collections now include their complete rows. Read report.result_rows or call to_python_objects(); accessing an existing report does not execute the query again. Repeated identical collections reuse the opened source and prepared operation while computing fresh results. Source replacement fails the current call; a later explicit call can prepare the changed source.

workflow = (
    ctx.read_vortex("cargo.vortex")
    .select(["cargo_id", "load_units"])
    .drop_duplicates(subset=["cargo_id"], keep="last")
)
report = workflow.collect()
print(report.result_rows)
workflow.write_vortex("deduplicated-cargo.vortex")

These collections admit at most 65,536 rows, 128 scalar fields and 8 MiB of JSONL. Larger results use bounded batches through write_vortex, write_parquet, write_arrow_ipc, write_avro, write_orc, write_json, write_jsonl or write_csv, subject to each format's dtype contract and the operation's state budget. Native Vortex and text output admit supported mixed scalar melt/pivot results; mixed Variant columns are not general binary compatibility output support.

Current source builds compose flat-scalar DISTINCT, drop_duplicates, duplicated, tail, sample, scalar rewrites, melt and rolling with admitted filters, projections, ordering, aggregates, joins and sets. Each operation consumes the preceding stage's rows and column names. SQL and DataFrame calls carry the same source declarations, CPU/memory allocation and writer policy through one native execution. Melt can omit ID columns; inferred value columns come from the preceding output. See the composition contract and acceptance. Static List/FixedSizeList/Struct payloads compose through admitted relational stages and ordered/repeated explode. Nested payloads can be written as Vortex, JSON, JSONL, Arrow IPC, Parquet or Avro when the dtype is representable; nested CSV translates nested values to quoted JSON text cells without preserving their logical dtype; ORC rejects nested output. Exploded flat output supports all eight writers. Current source builds extend relational keys to static List, FixedSizeList and Struct values for equality, hashing and ordering in the existing join, set, group, sort, window and subquery kernels. COUNT, COUNT DISTINCT, MIN/MAX, comparisons, NULL tests and selected CASE/COALESCE/NULLIF results use the same native nested values. Key schemas must match recursively, including field names and order, list kind and width, leaf widths, decimal precision/scale and temporal identity; recursive nullability does not affect key compatibility. Retained nested values are admitted for DISTINCT/duplicate selection and masks, tail, sampling, parent-level forward fill (a NULL parent takes the prior complete value; child NULLs do not trigger filling), lossless same-shape melt and rolling COUNT. See the nested key and retained-state contract; it is merged with complete local and hosted check evidence. General Variant/extension operations, nested arithmetic/string operations, wider analytic-window behavior, adapters and general state spill remain outside this scope. Scalar pivot type/domain restrictions remain in force. See also the nested payload contract. Binary, Decimal128 (precision 1–38, scale 0–precision), Date32 and timezone-free microsecond timestamps can travel as payloads, including nested leaves. Their flat equality, hashing and ordering are admitted for relational joins, sets, groups, windows and subqueries, with COUNT/COUNT DISTINCT/MIN/MAX and scoped comparisons and expressions; Decimal key precision and scale must match. Nested keys follow the subsequent contract above. Current source builds admit typed literals, explicit CAST/TRY_CAST, exact decimal arithmetic/rounding and scoped binary/calendar functions through the shared native expression binder. Decimal arithmetic output metadata binds before execution; explicit decimal downscaling requires zero discarded digits. Key compatibility still requires matching decimal precision/scale and preserves distinct temporal types. Wider analytic-window semantics, broader adapters and state spill remain separate. See the typed expression contract. Flat typed values also retain exact logical types through duplicate selection and masks, tail/sample, replacement/forward-fill, lossless melt, rolling COUNT and scoped pivot first/first-unique/COUNT. Python bytes, Decimal, date and datetime declare exact native literals. Decimal rewrites reuse checked native arithmetic; primitive predicate and typed sampling-weight restrictions remain. See the typed unary contract. Computed aggregate arguments and exact decimal aggregate, rolling and scalar pivot reductions now have complete local acceptance through the shared native engine. SUM uses precision 38 at the input scale; AVG uses precision 38 at max(input_scale,6) and rejects inexact division. Final overflow fails explicitly. ARRAY/STRUCT constructors preserve admitted logical child types. The revised engine report records 20,445 public checks, 202 direct checks and all 129 Full43 executions, with no benchmark-specific executor or external fallback. See the typed key contract. Binary supports all eight writers; ORC rejects decimal and temporal payloads. JSON/JSONL and collection encode binary as lowercase hex, decimals as decimal128(precision,scale):unscaled_integer, and temporal values as signed integer units. CSV uses the existing JSON scalar cell convention for binary/decimal strings. Text output does not retain native logical types. See the typed payload contract.

Current source builds also compose scalar pivot and pivot_table at their declared position, including renamed/ordered input, downstream filters, projections, aggregates, joins, sets, windows, melt and successive pivots. SQL uses PIVOT((SELECT ...), '{"index":"entity","columns":"category","values":"amount","aggregate":"sum"}'). Observed domains determine the columns during native execution; preparation and inspection do not read rows to guess them. Empty input keeps its actual index-only schema, with the declared margins column when requested. Referencing an absent domain fails explicitly. Correlated inner pivots bind independently for each outer row. All eight writers accept representable scalar results above the small collection limit, subject to the existing 128-field and memory limits. Pivot state has no spill path. See the dynamic pivot contract and acceptance.

For explicit preparation, use LazyFrame.prepare_vortex(...) or ctx.prepare_vortex(source_path, target_path, ...) with their documented arguments. These are preparation APIs; ordinary query execution uses the shared workflow below.

Engine intent is explicit. engine="auto" selects the current bounded snapshot batch path when allowed; live selects the CG-22 in-memory fixture path for bounded/unbounded change streams; hybrid selects the CG-22 declared Vortex-base plus in-memory hot-delta fixture for snapshot/bounded base overlays:

import shardloom as sl

ctx = sl.context(engine="live")
selection = ctx.engine_selection(
    boundedness="unbounded",
    update_mode="append-only",
    output_mode="changelog",
)
matrix = ctx.engine_capability_matrix()

print(ctx.engine)
print(selection.selection_status)
print(selection.selected_engine_mode)
print(selection.rejection_reasons)
print(matrix.engine_modes)
print(matrix.live_hybrid_claim_blocked_count)
print(matrix.live_hybrid_fabric_gate_rows)
print(matrix.live_hybrid_fabric_gate_claim_gate_status)
print(matrix.live_hybrid_fabric_gate_no_fallback_no_external_engine)

These calls do not execute workloads, probe brokers, write checkpoints, invoke external engines, or attempt fallback. They expose the same CG-22 contract as shardloom engine-selection-plan, shardloom engine-capability-matrix, and shardloom capabilities engines.

ctx.engine_capability_matrix() also exposes the GAR-0034-A live/hybrid fabric freshness gate. The gate keeps broker, state-store, object-store, catalog, production freshness, and exactly-once claims blocked unless future workload-scoped evidence promotes them, while preserving fallback_attempted=false, external_engine_invoked=false, and claim_gate_status=not_claim_grade.

The executable live surface is intentionally narrower: a deterministic in-memory fixture for filter, project, count, count_where, and group_count. It does not read brokers or files and does not write checkpoints, but it does emit freshness, state, continuous-view, execution, and Native I/O certificate fields:

contract = ctx.live_change_contract_plan()
fixture = ctx.live_fixture_run("group-count", "metric")

print(contract.change_record_fields)
print(contract.operations)
print(fixture.output_rows)
print(fixture.all_certified)
print(fixture.fallback_attempted)

Equivalent CLI commands:

shardloom live-change-contract-plan --format json
shardloom live-fixture-run group-count metric --format json

The executable hybrid surface is also fixture-scoped. It merges declared local Vortex base rows with deterministic hot deltas, applies tombstones/deletion vectors in memory, and emits delta-overlay, hot/cold contribution, micro-segment flush, layout-health, freshness, execution, and Native I/O evidence without reading or writing data:

hybrid = sl.context(engine="hybrid").hybrid_overlay_run("group-count", "metric")

print(hybrid.output_rows)
print(hybrid.layout_health_status)
print(hybrid.all_certified)
print(hybrid.write_io)

Equivalent CLI command:

shardloom hybrid-overlay-run group-count metric --format json

The first CG-23 REST/API surface is contract-first. It checks the versioned OpenAPI /v1 contract and the discovery-mode serve contract without starting a server, opening a listener, probing datasets, touching object stores, or executing queries:

api = ctx.rest_api_contract_plan()
discovery = ctx.serve_discovery_contract()
preview = ctx.rest_api_plan_preview("certified-local-batch")
lifecycle = ctx.rest_api_local_lifecycle("certified-local-batch")
events = ctx.rest_api_event_stream("certified-live-fixture")
security = ctx.rest_api_security_governance("safe-local-default")
data_plane = ctx.rest_api_data_plane("standards-matrix")

print(api.openapi_contract_path)
print(api.represented_resources)
print(api.discovery_endpoint_paths)
print(api.rest_runtime_unsupported_rows)
print(api.rest_runtime_unsupported_claim_gate_status)
print(api.rest_runtime_no_server_no_fallback_no_external_engine)
print(api.server_started)
print(discovery.server_mode)
print(discovery.contract_only)
print(preview.plan_handle)
print(preview.stage_statuses)
print(preview.problem_details_emitted)
print(lifecycle.lifecycle_status)
print(lifecycle.result_ref)
print(lifecycle.result_policies)
print(lifecycle.arrow_ipc_materialization)
print(lifecycle.fallback_attempted)
print(events.event_stream_status)
print(events.delivery_protocols)
print(events.event_types)
print(events.asyncapi_contract_path)
print(events.broker_io)
print(security.governance_status)
print(security.auth_postures)
print(security.api_scopes)
print(security.mcp_tools)
print(security.evidence_model_signals)
print(security.secrets_redacted)
print(data_plane.transfer_modes)
print(data_plane.preferred_large_payload_modes)
print(data_plane.standards_names)
print(data_plane.flight_adbc_required_for_basic_local_use)

Equivalent CLI commands:

shardloom rest-api-contract-plan --format json
shardloom rest-api-plan-preview certified-local-batch --format json
shardloom rest-api-plan-preview unsupported-operator --format json
shardloom rest-api-local-lifecycle certified-local-batch --format json
shardloom rest-api-local-lifecycle blocked-uncertified --format json
shardloom rest-api-event-stream certified-live-fixture --format json
shardloom rest-api-event-stream broker-requested --format json
shardloom rest-api-security-governance safe-local-default --format json
shardloom rest-api-security-governance destructive-policy-required --format json
shardloom rest-api-security-governance agent-mcp-discovery --format json
shardloom rest-api-data-plane artifact-reference-default --format json
shardloom rest-api-data-plane flight-ticket-requested --format json
shardloom rest-api-data-plane adbc-endpoint-requested --format json
shardloom rest-api-data-plane standards-matrix --format json
shardloom serve --mode discovery --format json

The GAR-0035-A REST runtime unsupported gate keeps HTTP listener, remote execution, Flight/ADBC transport, external broker integration, and dependency-expanded server claims blocked. The REST contract remains a checked-in OpenAPI/reporting surface until separate workload, server lifecycle, security, Native I/O, execution-certificate, and no-fallback evidence exists.

Lazy workflow planning is also available without adding pandas, Polars, Spark, DataFusion, or any other execution dependency:

import shardloom as sl

ctx = sl.context()
workflow = (
    ctx.read_vortex("orders.vortex")
    .filter("gte:value:3")
    .select("order_id", "amount")
    .limit(10)
)

plan = workflow.plan()
explain = workflow.explain()
estimate = workflow.estimate()
certification = workflow.certify()
unsupported = workflow.unsupported_report()

print(workflow.operation_summary)
print(plan.field("plan_only"))
print(explain.status)
print(estimate.status)
print(certification.fallback_attempted)
print(unsupported.fallback_attempted)

The same top-level helpers are exported as sl.read_vortex, sl.read_csv, sl.read_json, sl.read_parquet, sl.read_arrow_ipc, sl.read_avro, and sl.read_orc. Most helper chains still declare sources and transformations only. plan(), explain(), estimate(), certify(), and unsupported_report() are explicit report calls over CLI JSON surfaces; they do not read input files, infer schemas, materialize rows, probe object stores, write output, or invoke fallback engines.

Public Local Runtime: Universal Ingest Into A Vortex Middle

ShardLoom's Python front door is format-neutral at the execution boundary. ctx.read(path) and the explicit ctx.read_csv(...), ctx.read_json(...), ctx.read_parquet(...), ctx.read_arrow_ipc(...), ctx.read_avro(...), ctx.read_orc(...), and ctx.read_vortex(...) helpers are input adapters. They do not create separate CSV, JSON, Parquet, Arrow, Avro, ORC, SQL, or DataFrame execution stacks.

Local compatibility inputs enter the shared ShardLoom workflow, which converts them to the common Vortex-native execution representation before native computation. Source format, SQL, and the Python query builder do not create separate execution engines. collect() returns the complete typed result for admitted work; write(...) runs that same workflow to the declared sink.

The invariant on admitted public local workflows is:

input adapter -> shared Vortex-native execution -> typed result or declared sink
fallback_attempted=false
external_engine_invoked=false

Feature-gated structured adapters such as Parquet, Arrow IPC, Avro, and ORC still require the matching build/runtime feature. When an adapter, operator, or sink is not enabled or not admitted, the Python client returns the CLI unsupported envelope with a stable blocker id and next action. That is intentional: unsupported work fails closed instead of using pandas, Polars, DuckDB, Spark, DataFusion, or another engine.

A normal local Python use looks like this:

import shardloom as sl

ctx = sl.context()
orders = ctx.read("target/orders.csv")

result = (
    orders
    .filter(sl.col("amount") >= 10)
    .select("id", "amount", "status")
    .limit(10)
    .collect()
)

print(result.status)
print(result.result_rows)
print(result.fallback_attempted, result.external_engine_invoked)

The equivalent scoped SQL front door uses the same lifecycle after source parsing:

sql_result = ctx.sql(
    "SELECT id, amount, status FROM 'target/orders.csv' "
    "WHERE amount >= 10 LIMIT 10"
).collect()

print(sql_result.status, sql_result.result_rows)
print(sql_result.fallback_attempted, sql_result.external_engine_invoked)

Direct native Vortex input skips compatibility preparation and starts at the Vortex boundary:

vortex_result = (
    ctx.read_vortex("target/orders.vortex")
    .filter(sl.col("amount") >= 10)
    .select("id", "amount", "status")
    .limit(10)
    .collect()
)

print(vortex_result.status, vortex_result.result_rows)
print(vortex_result.fallback_attempted, vortex_result.external_engine_invoked)

Output adapters accept the selected workflow result when the input, operations, sink, and build features are admitted. Vortex remains the native persistence target; compatibility formats are explicit output translations. Unsupported combinations return a deterministic report. For example:

result = orders.write_jsonl("target/orders.jsonl", check=False)
print(result.output_path)
print(result.output_commit_status)
print(result.native_io_certificate_status)
print(result.fallback_attempted, result.external_engine_invoked)

Use write_json(...) when a single top-level JSON array is preferred:

result = orders.write_json("target/orders.json", check=False)
print(result.output_path)
print(result.output_commit_status)

Evidence-aware optimizer traces are planned as GAR-PERF-2B, not current Python runtime support. A future Python explain() trace should expose optimizer rule status, before/after plan digests, rewrite safety, evidence preservation, no-fallback fields, and claim gates without implying broad SQL/DataFrame execution or Polars/DataFusion optimizer parity.

Reusable I/O state and broad cross-format fanout are separate capability/evidence questions. The public Python query surface uses one shared native workflow for local files, typed-memory inputs, and source-free SQL. Input formats and output adapters meet at that workflow; no format selects an external execution engine. Capability reports describe their own scope and do not establish runtime support or performance evidence.

Unsupported workflow operations return explicit diagnostics. Wrapper-level operations such as an undeclared row UDF expose a blocker and the evidence needed to admit it:

import shardloom as sl

ctx = sl.context()
workflow = ctx.read_csv("events.csv").filter("amount > 0")
blocked = workflow.apply("row_udf", check=False)
print(blocked.blocker_id)
print(blocked.required_evidence)
print(blocked.suggested_next_action)
print(blocked.fallback_attempted, blocked.external_engine_invoked)

native_rejection = ctx.sql("SELECT CALL_API('https://example.invalid/score') AS score").collect(check=False)
print(native_rejection.status, native_rejection.diagnostics)

Unsupported reports above are generated through workflow-unsupported-plan and return status="unsupported" with fallback_attempted=false; admitted reports preserve the same no-fallback fields on their success evidence. The methods do not use pandas, pyarrow, or numpy as execution engines, parse SQL, execute unsupported DataFrame expressions, render broad notebook runtime output, invoke Foundry/model services, or use another engine as fallback. Valid pandas/Arrow inputs are treated as explicit materialized snapshots that lower to generated-source user rows, not as hidden external execution.

The DataFrame-style surface also has a typed method capability matrix. Use it when a wrapper, notebook, or agent needs to know which familiar method names are lazy declarations, which have scoped runtime-smoke support, which are unsupported diagnostics, and which evidence gates bound each method:

import shardloom as sl

ctx = sl.context()
matrix = ctx.capabilities().dataframe_method_matrix

print(matrix.row_order)
print(matrix.plan_only_methods)
print(matrix.unsupported_methods)
print(matrix.all_no_fallback_no_external_engine)

join = matrix.row("join")
print(join.support_status)
print(join.blocker_id)
print(join.required_evidence)
print(join.claim_boundary)

This matrix is still claim-safe, but its statuses should be read through the Vortex-middle contract. Local compatibility rows are not successful public runtime routes merely because a lower-level smoke command exists. Terminal methods are one of: side-effect-free lazy declarations, production-admitted local workflows that normalize through Vortex preparation or native Vortex input, source-free GeneratedSource local-output rows, internal smoke safeguards, or deterministic unsupported reports.

For local compatibility inputs, admitted collect(), count(), preview(), head(), take(), schema/data-quality summaries, bounded decoded materialization, local compatibility writes, compatibility fanout, quarantine sinks, and exact benchmark-family provider shapes enter the same Vortex-prepared/native route described above. Profile summaries and broad production/export semantics require an admitted route contract: native .vortex metadata profiles are admitted for base read/select/limit shapes, and scoped describe(...) lowers to that same metadata-first profile route. Transformed row profiling, pandas-style percentile/options summaries, and broad production profiling remain blocked until a native Vortex materialization/profile contract admits them. Alias rows such as project, where, groupby, order_by, sort_values, merge, nlargest, scoped tail, and scoped deterministic sample are useful only when their normalized operation shape maps to an admitted native route; otherwise they return the matching unsupported report before data is read.

Generated-source rows such as ctx.from_rows(...), ctx.range(...), ctx.sequence(...), and source-free SQL have their own local output contracts. Those are not proof that local CSV/JSON/etc. compatibility files can use direct decoded sinks as a product runtime.

schema_contract(...) and validate_schema(...) are bounded local schema evidence surfaces after Vortex preparation. They are not broad schema registry, table constraint manager, or object-store/ lakehouse enforcement surfaces. profile(...) is admitted for metadata-first native .vortex base read/select/limit profiles and otherwise returns deterministic blockers until a route-specific profile/materialization contract exists. It is not a hidden pandas/Polars profiler, resource tracer, performance claim, or production observability surface. quarantine(...) is admitted for bounded local checks and optional local sink replay evidence. It is not object-store/table quarantine, production remediation, or a broad data-governance engine.

When the question is broader than one DataFrame method, use the front-door parity matrix. It separates workflows that already lower SQL, Python, and DataFrame-style code to the same ShardLoom runtime path from the gaps that still block arbitrary SQL/Python/DataFrame flexibility and performance-equivalence claims. The scoped v1 boundary is owned by docs/architecture/v1-front-door-runtime-scope.md:

parity = ctx.front_door_parity_matrix()

print(parity.scoped_local_front_door_parity_supported)
print(parity.flexible_anything_claim_allowed)
print(parity.performance_equivalence_claim_allowed)
print(parity.row("local_file_filter_project_limit").shared_runtime_path)
print(parity.row("arbitrary_sql_python_dataframe_breadth").blocker_id)

Use the semantic surface matrix when the question is "which API/SQL semantic family is covered?" instead of "which route is selected?" It is the agent-facing companion to the human parity doc:

semantic = ctx.front_door_semantic_surface_matrix()

print(semantic.dataframe_subset_claim_statement)
print(semantic.sql_claim_statement)
print(semantic.pandas_compatible_claim_allowed)
print(semantic.ansi_sql_compliant_claim_allowed)
print(semantic.row("dataframe_materialization").claim_boundary)

Scoped local-file rows are admitted only when they normalize through Vortex preparation or start from native Vortex input and then match an admitted primitive/provider route. Generated-output rows remain separate source-free local-output contracts. Bounded schema/data-quality previews and decoded Python/pandas/Arrow/NumPy materialization are explicit gap rows until native Vortex-derived evidence and export/materialization contracts close them. General Vortex workflows, object-store/lakehouse/ table I/O, arbitrary SQL/Python/DataFrame breadth, and cross-front-door performance equivalence remain explicit gap rows until correctness, Native I/O, execution-certificate, no-fallback, and benchmark evidence closes them.

The v1 Vortex runtime scope is owned by docs/architecture/v1-vortex-runtime-scope.md. Use ctx.local_vortex_primitive_route_report() for the feature-gated local Vortex primitive route ids, CLI commands, materialization boundaries, and no-fallback evidence posture; broad object-store Vortex, table/catalog Vortex, generalized Source/Sink, and broad Vortex SQL/DataFrame support remain outside that scope. The v1 SourceState/prepared-state scope is owned by docs/architecture/v1-source-prepared-state-scope.md. Use ctx.source_prepared_state_scope_report() to inspect the UniversalIngress -> SourceState -> vortex_ingest -> VortexPreparedState route, the direct transient boundary, reuse/invalidation case ids, golden fixture refs, and required benchmark evidence fields. This report is local and claim-gated; it is not a global hidden cache, external cache service, object-store/table prepared-state reuse, broad non-local preparation, or performance claim.

Package, DataFrame, and notebook readiness are also exposed as a separate typed matrix so local install smoke is not confused with public package publication or broad runtime support:

readiness = ctx.dataframe_notebook_package_readiness()

print(readiness.schema_version)
print(readiness.local_install_smoke_supported)
print(readiness.package_publication_ready)
print(readiness.dataframe_runtime_supported)
print(readiness.notebook_runtime_supported)
print(readiness.all_rows_no_fallback_no_external_engine)

publication = readiness.row("public_package_publication")
print(publication.support_status)
print(publication.blocker_id)
print(publication.required_evidence)
print(publication.claim_boundary)

This readiness matrix is report-only capability posture. It does not publish to PyPI/TestPyPI/Conda/Homebrew, import notebook or DataFrame dependencies, render rich notebook output, execute broad DataFrame plans, call package repositories, or invoke external engines. Public package publication, broad DataFrame runtime, and notebook runtime remain blocked until release and execution evidence gates pass.

The CG-21 ETL workflow surface also has a compact typed matrix for current local workflow posture. Use it when a wrapper, notebook, or agent needs one place to show which user workflows are ready or smoke-supported, which APIs are report-only, and which production/runtime claims remain blocked:

matrix = ctx.etl_workflow_matrix()

print(matrix.schema_version)
print(matrix.supported_local_rows)
print(matrix.report_only_rows)
print(matrix.blocked_rows)
print(matrix.all_no_fallback_no_external_engine)

blocked = matrix.row("object_store_runtime")
print(blocked.status)
print(blocked.blocker_id)
print(blocked.claim_boundary)

This matrix is side-effect-free capability posture. It does not run production ETL, SQL/DataFrame execution, object-store/lakehouse runtime, Foundry runtime, external engine execution, or package publication, and it does not create performance or Spark-displacement claims.

GAR-0037-A adds a wrapper/connector implementation registry on the API-surface capability view. Use it when a client, adapter, agent, or public docs page needs to distinguish the current source-tree Python wrapper from planned or blocked ecosystem connectors:

caps = ctx.capabilities()
registry = caps.wrapper_connector_registry
# Or: registry = ctx.wrapper_connector_registry()

print(registry.schema_version)
print(registry.ready_local_count)
print(registry.report_only_count)
print(registry.blocked_count)
print(registry.all_rows_no_fallback_no_external_engine)

python = registry.row("python_cli_json_client")
sqlalchemy = registry.row("sqlalchemy")

print(python.support_status)
print(python.explicit_execution_available)
print(sqlalchemy.support_status)
print(sqlalchemy.deterministic_diagnostic_code)
print(sqlalchemy.claim_boundary)

The registry is capability posture, not connector implementation. It does not add generated clients, DB-API, SQLAlchemy, Ibis, dbt, Airflow, Dagster, Prefect, MCP, Flight SQL, ADBC, JDBC/ODBC, BI, Grafana, Foundry package, REST server, dependency expansion, network listener, external engine execution, or fallback. Rows preserve fallback_attempted=false, external_engine_invoked=false, and claim_gate_status=not_claim_grade.

Source-free generated-output APIs are tracked under GAR-GEN-1. The full contract is exposed through capability views as generated_source_contract, and the per-API admission matrix is exposed as generated_source_api_admission. GAR-NOVEL-1A also exposes generated_source_evidence_alignment, which ties the same GeneratedSourceCertificate rows to report-only OpenLineage, OpenTelemetry, Bayesian-confidence, and Foundry generated-output boundary refs without enabling exporters or platform runtime. Scoped local JSONL/CSV smoke paths are runtime-supported for caller-provided rows, Python literal tables, Python calendar/date dimensions, ShardLoom-native range/sequence generators, SQL VALUES, SQL literal SELECT, and SQL generate_series/range. Broader SQL/DataFrame forms remain report-only unless a later evidence-backed slice admits them:

caps = ctx.capabilities()
contract = caps.python.generated_source_contract
admission = caps.python.generated_source_api_admission
alignment = caps.python.generated_source_evidence_alignment
lineage = ctx.observability().openlineage_facet_mapping
telemetry = ctx.observability().opentelemetry_trace_export_contract

print(contract.schema_version)
print(contract.case_order)
print(contract.no_dataset_smoke_separate_from_generated_output)
print(contract.all_no_fallback_no_external_engine)
print(admission.row("python_ctx_from_rows").support_status)
print(admission.row("python_ctx_range").runtime_execution)
print(admission.row("python_ctx_sequence").runtime_execution)
print(admission.row("python_ctx_literal_table").support_status)
print(admission.row("python_ctx_calendar").runtime_execution)
print(admission.row("sql_values").support_status)
print(admission.row("sql_literal_select").runtime_execution)
print(admission.row("sql_generate_series_range").runtime_execution)
print(admission.all_no_fallback_no_external_engine)
print(alignment.schema_version)
print(alignment.openlineage_export_enabled)
print(alignment.opentelemetry_network_exporter_enabled)
print(alignment.row("foundry_generated_output").foundry_boundary_ref)
print(lineage.schema_version)
print(lineage.row("generated_source").facet_name)
print(lineage.all_rows_report_only)
print(lineage.all_no_fallback_no_external_engine)
print(telemetry.schema_version)
print(telemetry.row("operator_compute").timing_fields)
print(telemetry.network_exporter_enabled)
print(telemetry.no_export_side_effects)

The universal compatibility view also projects the same source-free generated-output posture so callers do not need to join GAR-GEN docs by hand:

compatibility = ctx.compatibility_scoreboard()
generated = compatibility.source_free_generated_output_contract

print(generated.schema_version)
print(generated.no_dataset_smoke_separate)
print(generated.local_output_only)
print(generated.output_certificate_required)
print(generated.row("python_ctx_from_rows").support_status)
print(generated.row("sql_values").support_status)
print(generated.row("local_output_only_generated_source_posture").blocker_id)
print(generated.all_no_fallback_no_external_engine)

This compatibility contract is a capability map, not a runtime report. Its generated-output rows describe declared posture only. Use the shared workflow execution report to inspect a real collect() or write request; unsupported shapes return deterministic diagnostics.

Source-free constructors such as ctx.from_rows(...), ctx.literal_table(...), ctx.range(...), ctx.sequence(...), and ctx.sql_values(...) return the same LazyFrame used by file-backed queries. Use .collect() for the complete typed result or a write_*() method for an admitted local sink. For example:

import shardloom as sl

ctx = sl.context()
frame = ctx.from_rows([{"id": 1, "label": "alpha"}, {"id": 2, "label": "beta"}])
collected = frame.collect()
written = frame.write_jsonl("target/generated-reference.jsonl")

print(collected.status, collected.result_rows)
print(written.status, written.output_path, written.output_commit_status)
print(written.native_io_certificate_status)
print(written.fallback_attempted, written.external_engine_invoked)

Source-free SQL expressions, VALUES, and range constructors use the same query path. A generated input does not imply support for every operation or sink: dtype, operation, adapter, and build feature admission still apply. The generated_source_* capability views are declarative maps; they do not constitute execution reports. Consult the live workflow report for status, and sink reports for output_commit_status and native_io_certificate_status.

The client also exposes the P7 claim gate closeout report:

from shardloom import ShardLoomClient

client = ShardLoomClient.from_repo()
closeout = client.claim_gate_closeout()

print(closeout.claim_gate_status)
print(closeout.release_readiness_status)
print(closeout.allowed_claims)
print(closeout.blocked_claims)
print(closeout.out_of_scope_claims)
print(closeout.no_runtime, closeout.no_fallback, closeout.no_effects)

This maps to shardloom claim-gate-closeout --format json. It is report-only: it does not run workloads, publish packages, probe APIs, run benchmarks, invoke Foundry, or permit external-engine fallback.

For P7.4 compute-engine closeout, the client exposes the report-only compute capability matrix and operator-family ladder:

from shardloom import ShardLoomClient

client = ShardLoomClient.from_repo()
matrix = client.compute_capability_matrix()

for row in matrix.rows:
    print(row.row_id, row.support_status, row.provider_kind, row.blocker_id)

for family in matrix.operator_families:
    print(family.family_id, family.support_status, family.next_evidence)

print(matrix.matrix_status)
print(matrix.claim_grade_status)
print(matrix.no_runtime, matrix.no_fallback, matrix.no_effects)

This maps to shardloom compute-capability-matrix --format json. It performs no runtime execution, data reads, writes, benchmark execution, external effects, external engine invocation, or fallback execution.

The first ShardLoomNative semantic conformance surface is executable, but only over side-effect-free in-memory fixtures. It records passed, planned, and blocked semantic dimensions before any broad SQL/DataFrame runtime claims:

from shardloom import ShardLoomClient

client = ShardLoomClient.from_repo()
suite = client.semantic_conformance_suite()

print(suite.semantic_profile)
print(suite.suite_status)
print(suite.executed_fixture_count, suite.passed_fixture_count)

for row in suite.rows:
    print(row.row_id, row.fixture_status, row.blocker_id)

This maps to shardloom semantic-conformance-suite --format json. Current fixtures cover the supported in-memory semantic dimensions and keep external oracles, dataset reads, SQL parsing, runtime execution, writes, and fallback disabled.

Artifact-rich top-level execution result envelopes can be inspected with ExecutionResultEnvelopeView when a command returns a shardloom.output.v2 execution envelope:

from shardloom import ExecutionResultEnvelopeView

def inspect_execution_envelope(envelope):
    result = ExecutionResultEnvelopeView(envelope)

    print(result.plan_id)
    print(result.provider_version)
    print(result.result_refs)
    print(result.artifact_refs)
    print(result.inline_artifact_ids)
    print(result.execution_certificate_refs)
    print(result.native_io_certificate_refs)
    print(result.representation_transitions)
    print(result.evidence_completeness_status)
    print([slot.kind for slot in result.incomplete_evidence_slots])
    print(result.fallback_attempted, result.external_engine_invoked)

The view is a typed reader over the CLI protocol. It does not execute unsupported work, create benchmark rows, write outputs, invoke external engines, or convert report-only surfaces into runtime support.

Package Build Smoke

The current source package version is owned by shardloom.__version__ and the shared workspace version sources. It is a Python client surface over ShardLoom's CLI, with bundled CLI resources in supported platform wheels. Bundled platform-wheel readiness can be checked locally without publishing:

python -m pip install build
python scripts/release_dry_run_proof.py --rows 64 --iterations 1

That proof builds the CLI, stages it under shardloom/bin/<system-arch>/ in a temporary package tree, builds a platform-specific wheel/sdist, installs the wheel in a clean environment, and asserts that ShardLoomClient().binary_command() resolves the bundled CLI without SHARDLOOM_BIN or SHARDLOOM_REPO_ROOT.

For a client-only wheel smoke without bundled CLI proof:

python -m build python
python -m venv $env:TEMP\shardloom-wheel-smoke
$wheel = Get-ChildItem python\dist\shardloom-*.whl | Select-Object -First 1
& $env:TEMP\shardloom-wheel-smoke\Scripts\python -m pip install $wheel.FullName
& $env:TEMP\shardloom-wheel-smoke\Scripts\python -c "import shardloom; print(shardloom.__version__)"

Conda packaging should stay split so the pure Python wrapper can remain noarch: python while the Rust CLI binary is built as a platform-specific package. Local recipe scaffolds live under packaging/conda/:

• shardloom-cli: compiled Rust shardloom binary.
• shardloom-python: pure Python wrapper/import surface.
• Optional shardloom metapackage: depends on both the wrapper and CLI for a one-command install path.

The recipes are not published packages. A release pass must align versions, replace local sources with tagged source archives and hashes, review license metadata, build packages in clean Conda environments, and receive explicit human approval before publication.

Spark, DataFusion, Polars, DuckDB, pandas, and Dask belong only in optional benchmark environments; they are not ShardLoom runtime dependencies or fallback engines.

Local Analytics Benchmark

Use the parameterized benchmark harness at benchmarks/traditional_analytics/run.py for local analytics comparisons. The example wrapper in examples/local-vortex-benchmark/ forwards its arguments to that harness and selects the public ShardLoom workflow. Supply a built CLI binary and workspace explicitly; the harness owns fixture generation, resource limits, and output handling. Benchmark results are workload- and host-specific evidence, not a general performance claim.

For application queries, use ctx.read(...) or a source-free constructor, then call collect() or write to a declared output with the same LazyFrame. See docs/getting-started/examples.md for runnable SQL and Python examples.

Quickstart Proof

The quickstart proof script stitches the local user flow together: import and CLI smoke, capability discovery, lazy source planning, unsupported explain/estimate diagnostics, compatibility-source planning, workflow readiness, and optional certified local Vortex primitive execution.

$env:PYTHONPATH = "python\src"
python python\examples\quickstart_proof.py --repo-root .

To include the currently certified fixture execution path, build the CLI with the local primitive feature and opt in explicitly:

python scripts\write_ci_version_env.py --format powershell | Invoke-Expression
$env:RUSTUP_TOOLCHAIN = $env:SHARDLOOM_RUST_MSRV_TOOLCHAIN
cargo build -p shardloom-cli --features vortex-local-primitives --bin shardloom

$env:PYTHONPATH = "python\src"
python python\examples\quickstart_proof.py --repo-root . --run-local-vortex

The optional execution path runs only the checked-in local_primitive_struct_five.vortex fixture through explicit local Vortex primitive flags. The planning portions remain no-write/no-probe, and the script exits nonzero if fallback is attempted, planning writes occur, or requested local primitive evidence is not certified.

Universal I/O is broader than local compatibility files. The current adapter registry also makes object-store, catalog, effectful, and unstructured queues visible from Python:

adapters = client.input_adapters()
print(adapters.field("common_structured_adapter_order"))
print(adapters.field("critical_structured_adapter_order"))
print(adapters.field("object_store_adapter_order"))
print(adapters.field("catalog_adapter_order"))
print(adapters.field("database_adapter_order"))
print(adapters.field("parquet_status"))
print(adapters.field("sqlite_status"))

plan = client.input_plan("file://tmp/example.parquet")
print(plan.field("source_kind"))
print(plan.field("capability_status"))
print(plan.field("plan_only"))

Common structured inputs are tracked as native_vortex, parquet, arrow_ipc, csv, JSON/NDJSON through jsonl, avro, and orc. Database adapters are visible separately: SQLite has a local import/export fixture smoke, while Postgres/MySQL, JDBC/ODBC, Snowflake, BigQuery, and Databricks SQL remain credential/network-gated. Lakehouse/table, object-store, catalog, effectful, and unstructured/media families are also represented in the registry. The current implemented live paths are scoped local fixture/evidence paths only: feature-gated local compatibility-file-to-Vortex benchmark smokes, native .vortex replay, public/local object-store fixture smokes, local table commit rehearsal, local SQLite import/export smoke, and the built-in deterministic scalar UDF fixture. Production adapter certification, live object-store runtime, catalogs, broad SQL/DataFrame runtime, arbitrary UDFs, and network connectors remain future work.

For a single source/sink compatibility view, use the typed scoreboard instead of scraping architecture prose:

matrix = ctx.compatibility_scoreboard()
print(matrix.schema_version)
print(matrix.row("vortex").support_status)
print(matrix.row("object_store_s3_gcs_adls").support_status)
print(matrix.all_rows_no_fallback_no_external_engine)

object_store = matrix.object_store_admission_ladder
print(object_store.schema_version)
print(object_store.provider_scope)
print(object_store.runtime_supported)
for row in object_store.rows:
    print(
        row.row_id,
        row.support_status,
        row.credential_policy_status,
        row.no_effects_no_fallback,
    )

The scoreboard maps local files, Vortex, generated outputs, Python rows, SQL literals, databases, object stores, table/lakehouse formats, remote APIs, and Foundry to runtime-supported, smoke-supported, report-only, blocked, or not-planned. It is a capability map only, not a production, performance, SQL/DataFrame, object-store/lakehouse, Foundry, or package claim. The object_store_admission_ladder keeps S3/GCS/ADLS URI recognition, credential policy, public reads, authenticated reads, byte-range reads, full-file reads, local cache, write staging, and commit protocol as separate gates. Current rows keep credential resolution, provider probes, network probes, object-store I/O, writes, commits, external engines, and fallback disabled. Important row IDs include object_store_uri_parse, credential_policy, public_no_credential_read, authenticated_read, byte_range_read, full_file_read, local_cache, write_staging, and commit_protocol.

For the first explicit object-store read runtime proof, use the local-emulator smoke. It reads a local fixture file through an object-store-style profile and emits SourceState, byte-range/full-file read, Native I/O, and no-fallback evidence.

read = client.object_store_read_smoke(
    "target/object-store-fixture.bin",
    byte_range=(0, 16),
)
print(read.field("object_store_read_status"))
print(read.field("source_state_id"))
print(read.field_bool("network_probe_performed"))
print(read.field_bool("fallback_attempted"))

For the public no-credential fixture profile, pass a supported S3/GCS/ADLS URI and an explicit local fixture file. ShardLoom parses the provider URI and reads the fixture bytes only; it does not resolve credentials, probe the provider, or open a network connection.

public_read = client.object_store_read_smoke(
    "s3://shardloom-public-fixtures/orders.vortex",
    profile="public-no-credential-fixture",
    public_fixture_path="target/object-store-public-fixture.vortex",
    fixture_listing=True,
    byte_range=(0, 16),
)
print(public_read.field("object_store_uri_parse_status"))
print(public_read.field("native_io_certificate_status"))
print(public_read.field_bool("public_no_credential_fixture_claim_allowed"))
print(public_read.field_bool("network_probe_performed"))

For the first explicit object-store write runtime proof, use the separate local-emulator write smoke. It stages a local source file into a local-emulator target path, commits a sidecar manifest, emits idempotency and digest evidence, and can immediately roll back the object plus manifest for cleanup proof.

write = client.object_store_write_smoke(
    "target/source.bin",
    "target/object-store-fixture.bin",
    idempotency_key="orders-batch-001",
    rollback_after_commit=True,
)
print(write.field("object_store_write_status"))
print(write.field("commit_protocol_status"))
print(write.field("rollback_status"))
print(write.field_bool("object_store_write_io"))
print(write.field_bool("fallback_attempted"))

Object-store read/write smokes remain fixture-scoped. Live real S3/GCS/ADLS network reads, credentials, provider probes, signed URLs, authenticated cloud reads or writes, cache writes, table/lakehouse commits, catalog interaction, distributed runtime, and production object-store claims remain blocked.

For the local SQLite adapter fixture, create or point at a local SQLite file and use the import/export smoke. The command table-scans a named table, writes a workspace-safe JSONL export, and creates a roundtrip SQLite artifact. It does not accept arbitrary SQL, push queries down, connect to network databases, resolve credentials, load extensions, or use SQLite as a fallback engine. order_by is post-scan fixture ordering in ShardLoom, and BLOB schemas/values are rejected.

sqlite = client.sqlite_local_import_export_smoke(
    "target/orders.sqlite",
    table="orders",
    export_jsonl="target/orders-sqlite.jsonl",
    roundtrip_db="target/orders-roundtrip.sqlite",
    order_by="id",
    allow_overwrite=True,
)
print(sqlite.field("sqlite_sql_execution_scope"))
print(sqlite.field_bool("sqlite_query_pushdown_allowed"))
print(sqlite.field("sqlite_ordering_execution_scope"))
print(sqlite.field_bool("roundtrip_replay_verified"))

For the built-in deterministic scalar UDF fixture, use the nullable-int64 fixture smoke. It proves UDF metadata, determinism, null propagation, overflow blocking, and effect policy for one built-in fixture only. It is not Python, WASM, Rust plugin, SQL-defined, table-function, or external-service UDF support.

registry = client.udf_registry()
print(registry.field("typed_udf_registry_support_status"))
print(registry.field_int("typed_udf_registry_admitted_local_fixture_count"))
print(registry.field_bool("typed_udf_registry_arbitrary_runtime_bridge_available"))

udf = client.udf_local_scalar_fixture_smoke([1, None, 3])
print(udf.field("udf_id"))
print(udf.field("output_values"))
print(udf.field_bool("external_effect_executed"))
print(udf.field_bool("fallback_attempted"))

Extension metadata and UDF runtime posture remain inspectable without executing extension code. A local extension manifest can be inspected as bounded metadata; the CLI does not load extension code, resolve credentials, probe networks, or enable plugin runtime support. The same helpers are available on ShardLoomContext when you want one high-level workflow surface:

extensions = client.extension_registry()
extension_dir = client.extension_registry(manifest_dir="target/extensions")
manifest = client.extension_inspect(manifest_path="target/extension.json")
typed_udfs = client.udf_registry()
fixture_plan = client.udf_runtime_plan("fixture")
python_plan = client.udf_runtime_plan("python")
print(extensions.field("extension_manifest_effect_all_runtime_blocked"))
print(extension_dir.field("extension_registry_manifest_count"))
print(extension_dir.field_bool("extension_registry_extension_code_executed"))
print(manifest.field("extension_manifest_inspection_status"))
print(manifest.field_bool("extension_manifest_execution_contract_complete"))
print(manifest.field_bool("extension_manifest_extension_code_executed"))
print(typed_udfs.field("typed_udf_registry_row_order"))
print(typed_udfs.field_bool("typed_udf_registry_external_engine_invoked"))
print(fixture_plan.field("udf_runtime_kind"))
print(python_plan.field_bool("udf_runtime_sandboxing_required"))

ctx_extensions = ctx.extension_registry()
ctx_extension_dir = ctx.extension_registry(manifest_dir="target/extensions")
ctx_manifest = ctx.extension_inspect(manifest_path="target/extension.json")
ctx_typed_udfs = ctx.udf_registry()
ctx_udf = ctx.udf_local_scalar_fixture_smoke([1, None, 3])
print(ctx_extensions.field_bool("extension_code_executed"))
print(ctx_extension_dir.field_bool("extension_registry_runtime_execution"))
print(ctx_manifest.field_bool("extension_manifest_external_effect_executed"))
print(ctx_typed_udfs.field_bool("typed_udf_registry_fallback_attempted"))
print(ctx_udf.field_bool("fallback_attempted"))

For the scoped local table metadata read proof, use the local-manifest smoke. It emits a typed metadata summary and digest evidence from ShardLoom's local manifest fixture without reading data files, touching object stores, resolving credentials, invoking table-format dependencies, or using fallback engines.

metadata = client.local_table_metadata_read_smoke()
print(metadata.field("support_status"))
print(metadata.field("claim_gate_status"))
print(metadata.field_bool("table_metadata_read_performed"))
print(metadata.field_bool("object_store_io_performed"))
print(metadata.field_bool("fallback_attempted"))

For the first fixture-scoped table append commit rehearsal, use the local manifest smoke. It writes a staged committed manifest plus sidecar table commit record, reports base/append/committed snapshot ids and digest evidence, and can immediately roll both artifacts back for cleanup proof.

table = client.local_table_append_commit_rehearsal_smoke(
    "target/table-commit/metadata-v2.json",
    idempotency_key="orders-table-commit-001",
    rollback_after_commit=True,
)
print(table.field("table_append_commit_status"))
print(table.field("committed_snapshot_id"))
print(table.field("commit_protocol_status"))
print(table.field_bool("table_catalog_commit_performed"))
print(table.field_bool("object_store_io"))
print(table.field_bool("fallback_attempted"))

The table metadata and append-commit smokes are local-manifest fixtures only. They are not Iceberg/Delta/Hudi production metadata/runtime support, catalog transactions, object-store-backed table commits, merge/update/delete runtime, distributed runtime, or performance claims.

The same scoreboard exposes table-format boundaries:

tables = matrix.table_format_boundary_matrix
print(tables.schema_version)
print(tables.format_scope)
print(tables.local_metadata_smoke_available)
print(tables.runtime_supported)
for row in tables.rows:
    print(row.row_id, row.support_status, row.no_io_no_fallback)

The table_format_boundary_matrix keeps Iceberg, Delta, and Hudi metadata reads, table scans, snapshot/time-travel, partition evolution, delete/tombstone, append, merge/update/delete, commit, rollback, catalog interaction, and object-store coupling as separate gates. Local manifest metadata, delete/ tombstone, and append commit rehearsal smokes are related evidence only; they are not production table-format runtime, lakehouse runtime, catalog runtime, object-store runtime, or commit support. Important row IDs include table_metadata_read, table_scan, snapshot_time_travel, partition_evolution, delete_tombstone, append, merge_update_delete, commit, rollback, catalog_interaction, and object_store_coupling.

The same scoreboard exposes database and warehouse import/export boundaries:

endpoints = matrix.database_warehouse_boundary_matrix
print(endpoints.schema_version)
print(endpoints.endpoint_scope)
print(endpoints.runtime_supported)
for row in endpoints.rows:
    print(
        row.row_id,
        row.support_status,
        row.credential_required,
        row.network_required,
        row.no_effects_no_fallback,
    )

The database_warehouse_boundary_matrix keeps SQLite, Postgres, MySQL, JDBC/ODBC, Snowflake, BigQuery, and Databricks SQL separate. SQLite is the only admitted fixture exception: sqlite_file is smoke-supported for local named table import/export through sqlite-local-import-export-smoke, with query pushdown disabled and no credentials/network probes. Postgres/MySQL, JDBC/ODBC, Snowflake, BigQuery, and Databricks SQL remain blocked as connectors and cannot serve as fallback engines. Important row IDs include sqlite_file, postgres, mysql, jdbc_odbc, snowflake, bigquery, and databricks_sql.

The client also exposes advisory optimization reports:

dynamic = client.dynamic_work_shaping_plan("memory-pressure")
sizing = client.sizing_feedback_plan(8, ["task-too-large", "memory-pressure-high"])

These commands report planned/advisory state only; they do not mutate runtime policy yet.

Planning and evidence commands may return status="success" while including error-severity diagnostics that describe missing evidence or blocked future work. The Python client preserves those diagnostics for inspection instead of raising unless the CLI exits nonzero or the envelope status is error or unsupported.

The example script wires the same calls together:

$env:PYTHONPATH = "python\src"
python examples\local-vortex-benchmark\run.py `
  --shardloom-binary target\release\shardloom.exe `
  --workspace "$HOME\LocalData\shardloom\traditional-benchmarks" `
  --input-state raw --output-format collect --reference-engine pandas

Test

$env:PYTHONPATH = "python\src"
python -m unittest discover python\tests