test(edifice): minimal offline regression test for the mission pipeline (pull → report → push)

[BluegReeno]

Summary

We just ran the full /edifice loop by hand (pull → improve → report → push) and it surfaced a live backend bug (#74) only at the last step. Nothing guards this pipeline in CI, so an Archon change to hal-mcp or the edifice mapping can silently break the report that field users depend on (IC Ingénieurs today; suivi_chantier prospects such as G. Baron next).

Add a minimal, offline, deterministic regression test that locks the pipeline's contract. Goal: a change that breaks pull→report→push turns a CI job red, not a field mission.

Companion: BluegReeno/bluegreen-marketplace#44 covers the client-side render half (build_context → DOCX). This issue covers the backend half. Land them together for full coverage.

Scope (MVP)

Two fixtures under tests/fixtures/edifice/, mirroring what the PWA stores — one diagnostic and one suivi_chantier mission, each with ≥1 note and ≥1 photo reference (tiny stub images, not real photos):

1. Backend contract — push_mission_context (guards fix(hal-mcp): push_mission_context upserts partial note rows — every push fails on project_id NOT NULL #74). Given seeded edifice_notes, a push of partial observations must: UPDATE by note_id (never insert/upsert), preserve project_id and untouched columns, report updated = rows actually affected, and report an unknown note_id as skipped/error (never a phantom insert). Idempotent: pushing twice changes nothing. → Best placed as a Deno test beside supabase/functions/hal-mcp/index.ts, ideally landed with the fix(hal-mcp): push_mission_context upserts partial note rows — every push fails on project_id NOT NULL #74 fix; add a deno test step to CI if none exists.

2. Pull/report contract — get_mission_with_assets → context shape. Assert the assembled mission+assets payload keeps the shape the edifice skill's build_context.py consumes, for both project_types: notes carry note_id + photos[], assessment maps to the diagnostic IE scale / suivi_chantier states, photo filenames resolve. This pins the backend↔skill contract so a shape drift is caught here rather than in the field.

Assertions are structural invariants, not bytes — e.g. "2 observations, each with a note_id and a resolvable photo", never a hash of a rendered .docx (the renderer re-encodes images, so bytes change every run — a byte-exact test would be permanently red and get disabled).

Non-goals / companion

• DOCX render smoke test (build_context.py → render_diagnostic.py / render_cr_visite.py → assert "opens, N disorders, M images"): those scripts live in the skill/plugin repo (bluegreen-marketplace), not here. Tracked as the companion BluegReeno/bluegreen-marketplace#44; out of scope for this repo.
• No improve (LLM) step in the test — represent the qualified mission as a fixed context.json fixture, not a live model call.
• No live Supabase in CI — mock the client or gate behind the existing live marker; the default pytest -m "not live" run must stay green offline.

Acceptance

• New tests + fixtures under tests/ and tests/fixtures/edifice/, green under uv run pytest tests/ -q -m "not live" (CI's python job then runs them on every PR with no extra wiring). If a Deno test is added for fix(hal-mcp): push_mission_context upserts partial note rows — every push fails on project_id NOT NULL #74, it runs under a new CI deno test step.
• The push test would have caught fix(hal-mcp): push_mission_context upserts partial note rows — every push fails on project_id NOT NULL #74 (upsert → NOT NULL project_id): it fails on insert/upsert, passes only on UPDATE-by-note_id, and asserts idempotency + honest updated count.
• Both diagnostic and suivi_chantier fixtures covered.
• Fully deterministic and offline; runs in a few seconds.

Maintenance

Fixtures are versioned beside the tests. When a behavior change is intentional, update the fixture / expected invariants in the same PR and review the diff (golden-file discipline). Keep the improve step as a fixture and anything needing real Supabase behind live, so the offline suite never rots into a permanently-red state nobody trusts.

Related

• fix(hal-mcp): push_mission_context upserts partial note rows — every push fails on project_id NOT NULL #74 — the push_mission_context UPDATE-vs-upsert fix this test guards (land the push test with it).
• BluegReeno/bluegreen-marketplace#44 — companion: client-side build_context → render DOCX smoke test. The two together = the full pull→report→push net.
• bluegreen-marketplace#31 (closed) — original skill-side report of the push bug.

---

Filed from a Cowork session that walked the full loop end-to-end. Intentionally left unlabeled / not queued for Archon — pick it up when priorities allow.