fix(run6.3): smoke-found bench harness fixes (7 schema/protocol bugs)

[egerev]

Summary

Run 6.3 manual smoke fixes. Running just bench --fixture small end-to-end against a live stack (with ANTHROPIC_API_KEY set) caught 7 real bugs in scripts/bench.py and scripts/bench_report.py that were invisible to the existing unit-test suite (which uses a synthetic SQLite schema with whatever names the queries expect — i.e. doesn't catch production-schema mismatches).

Result: smoke now passes end-to-end. exit 0, all 9 phases complete, bench_results/{date}-{sha}/ contains valid summary.md + manifest.json (env_summary masked) + extraction_runs/integrator_runs/pipeline_traces JSON dumps.

Bugs fixed (in commit order)

#	Commit	Bug	Symptom
1	c07c9e9	phase_bootstrap didn't send slug on POST /workspaces (required by WorkspaceCreate schema). Also: .env's ALAYA_DATABASE_URL=...localhost... interferes with docker compose up because compose's ${VAR:-default} interpolation reads .env, overriding the compose-internal default.	Phase 4: 400 validation error; Phase 2: migrations container fails to connect to 127.0.0.1:5432.
2	5783f99	from scripts.bench_ingest import ... inside functions failed because when python scripts/bench.py runs, sys.path[0] is scripts/, not the repo root.	Phase 5: ModuleNotFoundError: No module named 'scripts'.
3	2dd7e9f	bench_report.py SELECTed created_at from extraction_runs and integrator_runs, but those tables have started_at (no created_at column). Also SELECTed nonexistent cortex_cost_usd/crystallizer_cost_usd and entities_merged (the latter is in extraction_runs but bench_report queried integrator_runs).	Phase 8: column "entities_merged" does not exist.
4	6ad3ff1	extraction_runs.started_at is nullable and never populated by any code path — temporal filter started_at >= :ts excluded all rows. Workspace isolation per-bench-run already provides scoping.	Phase 8 succeeded but result: "empty_workspace".
5	3a5b656	bench_report.py queried claims, entities, predicates, but the actual table names are l2_claims, l1_entities, predicate_definitions.	Phase 8: relation "claims" does not exist.
6	6aed181	bench_report.py JOIN'd integrator_actions on integrator_run_id, but the actual column is run_id.	Phase 8: column ia.integrator_run_id does not exist.
7	abaa7bf	_serialize_row handled UUID and datetime, but Postgres NUMERIC columns return Decimal, which json.dumps rejects.	Phase 8: Object of type Decimal is not JSON serializable.
extra (in commit 1)	c07c9e9	phase_integrator only accepted HTTP 200/201 from POST /integrator-runs/trigger, but the router returns 202 Accepted.	Phase 7: FAIL: integrator trigger returned 202.

Why unit tests didn't catch any of this

tests/test_bench_report.py builds an in-memory SQLite database with synthetic CREATE TABLE statements that match whatever the queries expect — it doesn't validate against production schema. The only way to catch these was an actual end-to-end smoke run against Postgres + the real worker pipeline. Lesson for Run 6.4: add an integration-marked test that runs format_summary against the real postgres testcontainers schema (the same fixture used by tests/test_migration_009_forward_down.py).

Smoke evidence (after this PR)

$ just bench --fixture small  (with ANTHROPIC_API_KEY set)
[bench] Phase 1: preflight             — OK
[bench] Phase 2: docker compose up     — OK (compose env override)
[bench] Phase 3: API health            — healthy
[bench] Phase 4: bootstrap workspace   — Workspace created (slug provided)
[bench] Phase 5: ingest small.jsonl    — 5 events queued
[bench] Phase 6: drain                 — 5/5 terminal
[bench] Phase 7: integrator            — completed (202 accepted)
[bench] Phase 8: report                — manifest + summary + JSON dumps written
[bench] Phase 9: teardown              — services stopped

Manifest excerpt:

{
  "result": "complete",
  "exit_code": 0,
  "extraction_runs_count": 5,
  "integrator_runs_count": 1,
  "pipeline_traces_count": 13,
  "env_summary": { "ANTHROPIC_API_KEY": "set", ... all others "unset" },
  "model_versions": { all "unset" },
  "cache_hit_ratio_per_stage": { "cortex": 0.0, "crystallizer": 0.0, "integrator:*": null },
  "quality_proxies": { "claims_per_event_stddev": 0.0, "dedup_actions": 0, "run_failure_count": 0, ... }
}

grep -lE "ak_[a-zA-Z0-9_-]{16,}" bench_results/.../* returns empty — no API key leaked into any artifact.

Verification (gates)

uv run ruff check .              — clean
uv run ruff format --check .     — clean
uv run pyright                   — 0 errors
uv run pytest                    — 1158 passed, 1 skipped, 84 deselected
uv run python scripts/bench.py --help — works
just bench --fixture small       — exit 0, complete artifact (smoke evidence above)

Test plan

• CI green
• After merge: just bench --fixture medium shows non-zero cost_usd (medium fixture uses real Anthropic calls; cortex/crystallizer should generate cost)
• Optional: just bench --fixture medium --cache-warm-check — second-run cortex cache_hit_ratio > 0.5 (SC5)

Known limitations / follow-ups (not in scope of this PR)

• Test coverage gap: unit tests still use synthetic SQLite schema. A future PR should add an integration-marked test that runs format_summary against real testcontainers Postgres (with all tables migrated). Tracking idea: RUN6.3.FU.03 — bench_report integration test against live schema.
• extraction_runs.started_at never populated: this is a pre-existing schema/code-path inconsistency unrelated to bench. The bench-side fix is to scope by workspace_id only. A future cleanup could decide whether to populate started_at in ExtractionRunRepository.update_status (when transitioning to processing) or drop the column.
• Run-level cost shows $0.000000 on small fixture: likely because recalc_usage ran but adapter cost flow needs verification on a larger fixture. Not blocking — small is for smoke; medium/large will give real numbers.

🤖 Generated with Claude Code