feat(e0): refuse an unroutable MIME at the E0 gate instead of dead-lettering it (D104)

[fazpu]

The defect

The conversion router was consulted for the first time inside the convert worker. An upload whose MIME had no route was admitted, hashed, written to immutable raw storage, returned as a normal accepted-not-ready receipt, and only then dead-lettered through UnroutableMimeError → mark_version_failed → NonRetryableHandlerError. The caller discovered a terminal state by polling, having been told the upload was accepted.

The waste was knowable at admission — the route table was already in memory — and the caller cannot undo it: identical bytes are the D55 no-op and no API lets them request reprocessing. Recovery is operator work only (remember ops replay reopens the dead-lettered work row, after which a now-routable version converts normally). Each wrong guess costs a durable raw object and an operator ticket.

This is not a media defect. The mechanism keys on absence from the table, so it fired identically for audio, video, images, office documents and archives. The stock table is two text entries.

The placement is the decision

Three ingresses reach E0 without sharing a handler: HTTP POST /ingest, the local MCP ingest tool, and the connector sync worker — the latter two calling the composed port directly. A check on the HTTP handler would have left two of three paths still admitting bytes convert can only dead-letter, while looking fixed.

UploadIngestor is the one object all three write through, which is what the library boundary already requires: ingestion always writes through E0. The check sits in _guard_ingest beside the D74 guard; surfaces only render it.

Two details make that hold rather than merely sound right:

• routable_mimes is required. A default of "no check" would make the invariant as strong as every composer remembering to pass it. Every deployment has a route table (the settings default is the stock text one), so omission expresses only a mistake.
• Routability is decided before the D74 per-source query. Both orders are safe since neither writes bytes, but deciding it first avoids an admission query for a request that cannot be accepted and stops a forget-state error from masking a plain "we do not convert that".

Scope of the guarantee

No upload is admitted under a MIME this deployment cannot convert — not no unconvertible version is ever created. The gate reads the declared MIME; convert reads content_objects.mime, which is first-write-wins per content hash. They diverge only when the same bytes return under a different MIME whose first-seen reading has since become unrouted. Closing that costs a content-object lookup on every ingest; the trade is recorded as an open question in the analysis rather than paid or dismissed silently, and the worker's UnroutableMimeError still covers the case.

Surfaces

HTTP renders 415 with the accepted set. map_backend_error maps the typed error for local MCP and the 415 for remote MCP, so an agent gets unsupported_media_type with an actionable next step on both instead of internal_error locally and a flattened engine_client_error remotely. The SDK preserves the structured detail, trusting a code only at its bound status and path.

Breaking change

rememberstack.workers.UploadIngestor now takes a required routable_mimes. Pass frozenset(conversion_routes) where you build it. Deployments configured through REMEMBERSTACK_SELFHOST_CONVERSION_ROUTES need no change. Recorded in /docs/project-status.

Corpus

• Decision: decisions.md D104
• Design: plan/designs/e0_files_design.md §3
• Analysis: plan/analysis/unroutable_mime_preflight.md

Verification

make check clean — ruff, pyright, 1308 passed / 679 skipped.

Tests assert store.writes == 0 on refusal (the exception alone would not prove it — the defect was that bytes became durable), cover both E0 entry points, prove D74 is not consulted for an unroutable input, and lock the gate/router key-set equivalence. Two structural audits guard the shape: record_upload has exactly the two gated callers (counted, not set-collapsed), and no runtime module outside DocumentCatalog inserts document_versions.

Reviewed across five adversarial codex rounds; findings from every round are applied.

🤖 Generated with Claude Code

https://claude.ai/code/session_01KfBUpQuuehJKf2kbNa2D2D

Contributor agreement

Signing on behalf of a legal entity (leave blank if accepting individually):

• I have read and agree to the RememberStack Contributor License Agreement v1.0.