emit

Imported from morzecrew/bloomery (src/bloomery/emit/AGENTS.md). Install upstream with npx skills add morzecrew/bloomery --skill emit. Copyright stays with the author.

Decisions governing src/bloomery/emit/

S-0002/D-7 — ASSUMED (Multi-project composition) — implementation: partial

Two composing projects must share a naming policy; whether a mismatch is a refusal depends on the upstream IR recording the policy it was compiled under

• Paths: src/bloomery/naming.py src/bloomery/emit/base.py
• Consequence: Without a shared policy the downstream names relations the upstream never created, and the failure surfaces in the warehouse rather than in the compile — which is the worst place for it

S-0011/D-6 — ASSUMED (Retrieval semantics) — implementation: none

The retrieval manifest is emitted by a target of its own, not alongside another target's artifacts

• Paths: src/bloomery/compile.py src/bloomery/emit/**
• Consequence: A project's retrieval contract is independent of which analytical framework it compiles for, at the cost of a target enum member that names an artifact rather than a consumer

S-0011/D-7 — ASSUMED (Retrieval semantics) — implementation: none

The manifest inlines each profile's space fully rather than referencing it by name

• Paths: src/bloomery/emit/**
• Consequence: The manifest is larger and repeats a space shared by several profiles, and in exchange it is readable by a consumer that has no access to the spec tree

S-0011/D-10 — LOCKED (Retrieval semantics) — implementation: none

A vendor-oriented vector emitter is register_emitter, out of tree, and stays there until it has an artifact contract someone has run; no vector-database member of the target enum

• Paths: src/bloomery/compile.py src/bloomery/emit/__init__.py
• Consequence: The capability is available to anyone who wants it without core carrying a vendor's name or its compatibility promise
• Touching these paths owes a divergence entry: torve log owed <task> --touched <files> before you finish

S-0020/D-5 — ASSUMED (Intermediate representation and determinism contract)

Floats are banned in IR and emission; Decimal/int only.

• Paths: src/bloomery/cli/serialize.py src/bloomery/emit/lower/reconcile.py src/bloomery/emit/steps.py src/bloomery/ir/fingerprint.py src/bloomery/ir/nodes.py src/bloomery/planner/request.py src/bloomery/quality/predicates.py src/bloomery/resolve/steps.py src/bloomery/spec/common.py src/bloomery/spec/marts.py src/bloomery/spec/metrics.py src/bloomery/spec/quality.py src/bloomery/steps/manifest.py src/bloomery/steps/splice.py src/bloomery/transforms/_builtins.py src/bloomery/typing/types.py tests/execution/test_as_of_join.py tests/execution/test_currency_convert.py tests/execution/test_fanout_trap.py tests/execution/test_marts.py tests/execution/test_period_over_period.py tests/execution/test_rollup.py tests/fixtures/semantic_corpus/002-average-of-averages/naive.sql tests/golden/schema/entity_model.json tests/golden/schema/mapping.json tests/golden/schema/marts.json tests/support/planning.py tests/support/semantic_corpus.py tests/unit/test_cli.py tests/unit/test_emit/test_quality_mart.py tests/unit/test_planner/test_request.py tests/unit/test_quality/test_edges.py tests/unit/test_spec/test_metrics.py tests/unit/test_spec/test_quality.py tests/unit/test_steps/test_lowering.py tests/unit/test_steps/test_manifest_and_registry.py tests/unit/test_typing/test_types.py

S-0025/D-2 — ASSUMED (Ports and emitters: targets, dialects, naming)

Emitters produce file-shaped text artifacts as data (settles open question #1). No filesystem writes, no live-context registration in core — callers build that on top.

• Paths: src/bloomery/emit/base.py src/bloomery/emit/steps.py tests/golden/test_sqlmesh_duckdb.py tests/support/execution.py

S-0025/D-4 — ASSUMED (Ports and emitters: targets, dialects, naming)

SQL rendering: SQLGlot AST → DialectPort.render; Jinja only ever sees pre-rendered strings (envelope templating).

• Paths: src/bloomery/dialects/base.py src/bloomery/emit/dbt/__init__.py src/bloomery/emit/lower/predicates.py src/bloomery/emit/lower/silver.py src/bloomery/emit/sqlmesh/__init__.py src/bloomery/emit/steps.py tests/property/test_compile_properties.py tests/unit/test_steps/test_dbt_and_cube_emission.py

S-0025/D-8 — ASSUMED (Ports and emitters: targets, dialects, naming)

Emitter/dialect registries mirror the transform registry: immutable defaults + explicit overlay, collision is an error, iteration sorted (S-0021/D-6).

• Paths: src/bloomery/dialects/__init__.py src/bloomery/emit/__init__.py src/bloomery/runtime/sql_client.py tests/unit/test_dialects/test_base.py tests/unit/test_emit/test_base.py

S-0025/D-9 — ASSUMED (Ports and emitters: targets, dialects, naming)

Every artifact carries a header comment with the project fingerprint — applied-vs-spec drift detection downstream.

• Paths: src/bloomery/emit/base.py src/bloomery/emit/metricflow/__init__.py src/bloomery/emit/sqlmesh/__init__.py

S-0025/D-10 — ASSUMED (Ports and emitters: targets, dialects, naming)

(Amended for _bloomery-changes.md D6) TargetCapabilities is a frozenset[Feature] over a closed Feature StrEnum — membership-checked, sorted at any output-reaching iteration. The native planner (S-0028) is a fourth port sharing this vocabulary; its lack of QUERY_TIME_JOIN/MULTI_FACT is the fan-out-impossibility property, documented as a feature.

• Paths: src/bloomery/emit/base.py

S-0033/D-89 — ASSUMED (Data quality: declarative cleansing, dispositions, quarantine)

(2026-08-10) Mart-level checks are assertions, not quality rules. §10's open question is settled by the disposition model. §8 deferred them as "blurs into reconciliation" and §10 asked "reconcile-shaped or new surface?" — the answer is neither, and what decides it is not taste. §5.9 draws the boundary at what a verdict does: a quality rule disposes of a row. A mart row is derived — no _source_row_id, no bronze payload, no reject table, no replay — so there is nothing to quarantine, nothing to repair, and nothing to bring back; and a reconcile compares two sides, which "no month has zero revenue" is not. What is left is D4's other half, "alert me", so assert: on a mart declares {measure, agg, by, min/max, on_fail} and lowers to an audit the mart model names. quarantine and repair are absent from its on_fail rather than lowered to something weaker, so an author who wanted routing learns it at the surface instead of from a mart that silently only alerts; fail/flag map to blocking/non-blocking exactly as reconcile.on_fail does (D38). The aggregate vocabulary is deliberately the same tuple the reconcile grammar uses — both compute one number over a column so a human can be told it is wrong, and two lists that mean the same thing drift. Bounds ride in params as text through the D57 carrier, for the same reason. Resolution is against the flattened column set, not the base entity's: ordered_month exists only because a date: step made it, and it is precisely the column §10's example groups by. The body is SELECT <by…>, <agg>(<measure>) FROM @this_model [GROUP BY <by…>] HAVING <bound comparison> — the value beside the group, because a failure a human must open the warehouse to understand gets ignored. A bare HAVING with no GROUP BY is the whole-mart form; both shapes were executed on DuckDB, postgres 16 and trinodb/trino:483 rather than read out of three manuals, and all three agree. What it cannot see, stated rather than implied: D19 reaches the mart, so every aggregate but count is NULL over an empty group and the assertion stays silent — which is also why an assertion cannot notice a month that is entirely missing: no row means no group at all. count is the exception and is the one shape that catches an empty mart. Closing the missing-period case needs a join against the date spine, a coverage check with its own dependency, and it is named here rather than half-built. dbt refuses a project carrying one, sharing refuse_steps' message shape (amended by D94: this row said "dbt and Cube", and the Cube half was already obsolete when it was written — S-0034/D-52 had split the blanket Cube refusal on the argument that Cube builds no relation for anything, and Cube compiles the entire quality surface, quarantine and reject tables included, without a murmur; refusing this one check would single it out): neither has the audit form, and compiling clean while the declared gate does not exist is the failure D83 caught in the dialect ports. Cost recorded: MartIR gains a field, and the canonical encoder covers each node's field names and count, so every fingerprint in the corpus moves — the churn §12 budgets, spent deliberately. The fixture carrying the demonstration is quality_precedence rather than ecom_basic, because an assertion makes a project uncompilable for two targets and ecom_basic is the fixture those targets' goldens are built on.

• Paths: src/bloomery/emit/base.py src/bloomery/emit/dbt/__init__.py src/bloomery/emit/lower/reconcile.py src/bloomery/emit/sqlmesh/__init__.py src/bloomery/ir/nodes.py src/bloomery/marts/flatten.py src/bloomery/spec/marts.py tests/execution/test_quality_precedence.py tests/fixtures/quality_precedence/marts.yaml tests/golden/schema/marts.json

S-0034/D-8 — ASSUMED (The step registry: referenced implementations)

Emission: sql_macro splices into the entity SELECT (one query); sql_model emits an ordinary model artifact from the registry body; python_model emits a generated SQLMesh Python-model .py artifact (S-0025/D-2 file-shaped) wrapping the impl + contract assertion. Step outputs are DAG entities with manifest grain; two steps writing one output is a compile error (settles Document 5 §11.5).

• Paths: src/bloomery/emit/steps.py src/bloomery/resolve/steps.py tests/unit/test_steps/test_emission.py tests/unit/test_steps/test_lowering.py

S-0034/D-16 — ASSUMED (The step registry: referenced implementations)

Multi-output emission resolved — supersedes the draft §10 entry and its execute-exactly-once constraint (recorded honestly: that constraint is dropped, not satisfied): each declared output gets its own generated wrapper model, each executing the step and returning its own output; safe for correctly declared steps — nondeterministic steps are compile-refused and seeded steps re-execute with the same recorded seed (pure/seeded ⇒ identical results); residual risk recorded: a misdeclared step slips the compile check and, under N executions, can produce disagreeing sibling outputs within one run (behavioral gates catch run-to-run, not intra-run, divergence) — accepted for v1, with a cross-output consistency audit named as the demand-gated mitigation. assert_step_contract runs in every wrapper against all declared outputs, catching partial-output lies wherever the run starts. The N-executions-for-N-outputs cost is documented; a single-execution staging optimization is a demand-gated, named escape hatch — not built.

• Paths: src/bloomery/emit/steps.py src/bloomery/resolve/steps.py src/bloomery/steps/manifest.py tests/execution/test_step_consistency.py tests/golden/test_sqlmesh_duckdb.py tests/unit/test_steps/test_emission.py

S-0034/D-52 — ASSUMED (The step registry: referenced implementations)

(2026-08-10) D31's blanket refusal was one sentence covering two different targets, and it is right about neither in full. D31 refused steps on dbt and Cube because "their output relations would simply be missing" — checked per target, that argument splits. Cube builds nothing. It emits cubes and views over marts, and no silver model, no reject table, no replay statement and no audit for anything: the dirty_corpus fixture — twelve quality-carrying entities, twelve reject models, a conservation audit apiece — compiles to two files on Cube and always has, refusing none of it. So "the relation would be missing" was never a reason to refuse a step here; it is equally true of every silver entity, and singling steps out made this emitter refuse one build-side declaration among the many it already leaves to whoever maintains the tables. The refusal is removed, and its docstring now says what the contract actually is: Cube consumes tables SQLMesh maintains, and is deliberately silent about how. (The mart-assertion refusal added the day before, S-0033/D-89, is removed from Cube for the same reason and stays on dbt.) dbt builds, so a step must emit or refuse, and the answer is per tier. Tier 1 needs nothing: the splice happens at lowering, so a macro is already inside its consuming model on every target — which means Tier 1 worked on dbt throughout and D31 never claimed otherwise only because a macro is referenced inline rather than wired in steps:. Tier 2 emits: the body is already canonicalized and parameter-substituted on StepIR, so dbt wraps the same SELECT SQLMesh does in its own envelope — asserted as byte equality between the two targets rather than as a claim, since one SELECT meaning two things is exactly the drift the shared lowering exists to prevent. Tier 3 stays refused, on a concrete reason rather than a blanket one: dbt has Python models, but only on Snowflake, BigQuery and Databricks, and none of bloomery's three dialects is one of them, so the wrapper would have no adapter to execute it. Step audits are refused on dbt — a consistency audit (D40) is a join between sibling outputs and an on_fail: fail body (D39) is a whole query, while dbt's schema tests are per-column or per-model predicates; _entity_tests already refuses an audit kind on exactly this ground. The refusal is decided by building the audits rather than by the presence of a step, so a single-output Tier 2 step with no fail rules keeps the tier instead of losing it to a reason that does not apply to it. One trap, found and closed: a step output is an entity in the DAG (D36) whose lowered expr is the column referring to itself, so removing the refusal without skipping produced_by entities would have had dbt emit an ordinary entity model beside the step's — a model selecting from the relation it defines, and two models writing one relation. SQLMesh already skipped them; dbt now does too. Not verified, stated: that dbt parses the emitted model is S-0026's outstanding dbt parse tier, not something this row claims. What is verified is that the SELECT is byte-identical to SQLMesh's and that the envelope is the one the dbt goldens already lock.

• Paths: src/bloomery/emit/cube/__init__.py src/bloomery/emit/dbt/__init__.py src/bloomery/emit/steps.py tests/e2e/test_dbt_parse.py tests/golden/test_cube.py tests/unit/test_marts/test_asserts.py tests/unit/test_steps/test_dbt_and_cube_emission.py

S-0043/D-5 — LOCKED (The dbt singular-test surface)

The model reference goes through _reference_map, never string formatting. It is what makes a singular test a DAG participant rather than a query that happens to name a table, and it is already built for exactly this shape. Consequence: a singular test is ordered against its model by dbt, which is what makes "blocking" mean anything at all under D2.

• Paths: src/bloomery/emit/dbt/__init__.py src/bloomery/emit/lower/silver.py src/bloomery/emit/sqlmesh/__init__.py src/bloomery/emit/steps.py
• Touching these paths owes a divergence entry: torve log owed <task> --touched <files> before you finish

S-0043/D-10 — LOCKED (The dbt singular-test surface)

The audit envelope separates from the audit body (§5.3). emit/steps.py is target-neutral by position and not by content: its _AUDIT template bakes SQLMesh's AUDIT (name …); header in, so quality_audits and consistency_audits hand shared callers a SQLMesh artifact. A producer instead returns a named body — rendered SELECT, name, disposition — and each target wraps it. The alternative, a second template beside the first, gives one audit body two spellings maintained in parallel, which is the divergence the shared lowering exists to prevent. Consequence: this is the only structural change here, and it is what makes the remaining work a template and a path rather than five separate liftings. It also carries the relation in as a parameter, so @this_model stops being a literal the dbt side would rewrite by substitution.

• Paths: src/bloomery/emit/base.py src/bloomery/emit/dbt/__init__.py src/bloomery/emit/lower/silver.py src/bloomery/emit/steps.py tests/unit/test_emit/test_dbt.py tests/unit/test_steps/test_dbt_and_cube_emission.py
• Touching these paths owes a divergence entry: torve log owed <task> --touched <files> before you finish

Invariants holding over src/bloomery/emit/

• S-0002/I-1: A project carrying neither an exports nor an imports document compiles byte-identically to what it compiled before composition existed; the golden corpus is that test, and only a deliberate IR version bump may move a fingerprint in it
  • Paths: tests/golden/** src/bloomery/emit/**
  • Check: uv run pytest tests/golden -q