chore(b1): delete the /convert compatibility shim (PR C of 3, register #51) — DO NOT MERGE until N5 passes

[tornidomaroc-web]

PR C of 3 in the (b1) sequence. SERVICE-ONLY. DRAFT, and it must stay a draft until N5 passes.

Deletes the deprecated /convert compatibility shim from services/ingestion/main.py. That is the entire code change.

---

⛔ THIS MUST NOT MERGE UNTIL N5 PASSES

N5 (docs/b1-verification-protocol.md §3, §10) has two halves and both are required:

1. grep -rn "/convert" src/ returns zero hits, and
2. the production Vercel deployment's commit SHA is at or after PR B's merge commit.

Neither half is satisfied today. src/app/api/ingest/route.ts:154 still calls ${pythonServiceUrl}/convert, and PR B — the PR that repoints it — does not exist yet.

The second half is the one that cannot be shortcut. A merged PR B that Vercel has not finished deploying leaves production Next still calling /convert against an image that no longer serves it.

Merging this early is an instant production outage, not a degradation. Every upload's POST /convert returns 404, !pyResponse.ok fires, and every ingestion fails until a rebuild lands.

N6 is the post-merge backstop: if POST /convert does not return 404 after this merges, PR C did not deploy (§9 — Railway Watch Paths is empty today; a Watch Paths pattern set between B and C that never matches would let PR C merge green while /convert stays live forever, with no symptom anywhere).

---

⚠️ AND IT IS NOT OPTIONAL

Once PR B makes uploads work again, the temptation is to leave the shim in place as harmless. It is not.

If PR C is dropped:

• Register deps: bump @supabase/ssr 0.6.1 -> 0.10.2 (pinned exact) #50's 9-13× return-blob endpoint — full markdown plus every chunk's 1024-float embedding, returned inline — stays live indefinitely, behind a shared bearer token.
• The "loud 404 on version skew" property that the entire /convert → /ingest rename was chosen to buy is silently gone. A version-skewed caller would find a working old endpoint and quietly persist a null markdown_content over a document this service already finished correctly, instead of failing visibly.

The rename only pays for itself once the old name is actually absent. Until this PR lands, (b1) is not done — it is half-done with a regression parked in production.

---

Why this is stacked on PR A's branch, not on main

Based on main, this PR's diff would carry PR A's five files as well, and one merge would add /ingest and delete /convert simultaneously against the old caller — precisely the single-merge skew window the three-PR split was invented to eliminate, reachable from the green merge button.

Stacked on feat/b1a-service-ingest-with-convert-shim, the diff is exactly the two files below and the bad merge is mechanically unavailable. GitHub retargets this PR to main automatically when PR A merges, keeping this number, URL and body.

---

The diff, proven rather than asserted

git diff --numstat against PR A's head (c27163a):

4	0	docs/PROGRESS.md
0	48	services/ingestion/main.py

main.py is a pure deletion: 0 insertions, 48 deletions, the @app.post("/convert") handler and nothing else.

Survivors, from an ast walk of the post-deletion module — not from intent:

ROUTES:
  GET  /health  -> health
  POST /ingest  -> ingest
  POST /embed   -> embed_endpoint

FUNCS: _check_auth, _chunk_text, _convert_to_markdown, _embed, _embed_bge_m3,
       _embed_voyage, _load_owned_document, _mark_error, _persist,
       _safe_basename, _user_client, embed_endpoint, health, ingest

_convert_to_markdown is deliberately not deleted even though it sits directly above the shim and reads like part of it — it is the converter /ingest itself calls, so removing it with the handler would break the endpoint this whole sequence exists to ship.

No import is orphaned by the deletion, which is why there is no ride-along cleanup: UploadFile, File, Header, HTTPException, JSONResponse and Form are still used by /ingest, Body by /embed, and shutil / tempfile / MarkItDown by _convert_to_markdown. An unused-import removal here would have made the diff non-pure for zero benefit.

---

What this PR deliberately does NOT touch

• docs/b1-verification-protocol.md — untouched. The N5 and N6 result rows are left blank. §10 states plainly that an empty cell means the check has not been run and is not a pass by default; filling them at open time from a PR that has run neither check would be the exact vacuous-green failure register test(DO NOT MERGE): db-types drift probe #52 exists to end. They get filled when the checks actually run.
• Register row ci: generate database.types.ts from migrations and gate on drift #51 — untouched, and that is a defect report, not a preference. PR A introduced row ci: generate database.types.ts from migrations and gate on drift #51 at docs/PROGRESS.md:150 as a single-cell row in a four-column table (2 pipes, where every other row from Create test-sqli.js #1 to feat: wire <Database> into the three Supabase client factories #53 has 5 or more), so its Item, Why deferred and Where addressed cells are empty and its whole body renders inside the narrow # column. PR A is still open, so that structural defect belongs on PR A. Folding the repair in here would hide one PR's defect inside an unrelated one; updating the text while preserving the broken shape would carry it forward.
• Zero files under src/. No migration, no DB change, no requirements.txt change, no branch-protection change, no Railway or Vercel dashboard change. PIVOT_PLAN.md untouched, §1 B5b row stays ⬜, register docs: record live-DB reconciliation + "migration ≠ applied" lesson #17 stays OPEN, PR feat(b1): Railway ingestion service owns chunk persistence + terminal document status (forwarded user JWT, no service-role key) #70 left open.

---

Checks

Both pre-commit guards pass on the staged set:

secrets scan clean (2 files)
dash scan clean (0 user-facing files)

The required checks on this PR are not evidence. tsc and db-types will both pass vacuously — this PR contains zero TypeScript and zero SQL. The ingestion-image job (register #52) is the only check that reads a line of Python here, and it is not a required check.

---

Register #51. Sequence: PR A (#72, open) → PR B (does not exist yet) → PR C (this).

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
knowflow	Ready	Preview, Comment	Jul 28, 2026 3:28pm