scope FableLoom reformat to one episode per request so a large loom cannot time out

[atomantic]

Summary

Rewriting a loom's scenes into the other format ran every provider call behind a single held HTTP request — 227s on a 13-scene loom, and tens of minutes on a large one, long enough for a proxy or fetch timeout to kill the response while the server kept writing chunks to the record.

• The endpoint is episode-scoped: POST /api/fableloom/:id/episodes/:episodeId/reformat replaces the loom-wide POST /:id/reformat. Each request is additionally capped at 20 scenes; a run that stops at that ceiling with scenes it never sent answers capped: true.
• The settings drawer walks the episodes that still have work, names the one in flight ("Rewriting episode 2 of 3 — Pilot…"), and re-asks the same episode while it comes back capped. A response that is not capped is done with that episode even if the model dropped a scene — re-sending a refusal isn't progress, and the leftover rides the existing "run it again to finish" toast.
• The format pin stays the server's call, and lands only once no episode has an unconverted scene left. That is stronger than the final client PATCH the issue sketched: a browser closed mid-walk can't leave the loom claiming a format half its story isn't in.
• Resumability is untouched. Each chunk of 5 scenes is still persisted as it lands and stamped with the format it was written in, so a failure keeps what already succeeded and a re-run rewrites only what is left.
• The rewrite button now counts, and offers itself for, only the scenes not already in the target format — a fully converted loom no longer offers a pass that would do nothing.
• docs/features/fableloom.md ("Scene format") describes the new transport.

Option A from the issue. Option B (start + SSE + cancel, the Writers Room polish shape) stays available if reformat later grows a per-scene retry or starts rewriting image prompts too; nothing here blocks it.

Test plan

• server/services/fableLoom/weave.test.js — reworked for the episode-scoped service, plus new coverage: stops at the ceiling and continues where it stopped (capped: true → next request sends only the leftover chunk); does not flag a run as capped when the model, not the ceiling, left scenes behind; holds the loom pin until every episode is converted, not just the one it rewrote; is a no-op with zero provider calls on an already-converted episode; 404s on an episode not in the loom.
• server/routes/fableLoom.test.js — the reformat route validates the format and forwards (loomId, episodeId, input); the old loom-scoped path is gone (404).
• client/src/components/fableloom/LoomSettingsDrawer.test.jsx (new) — walks episodes one request at a time, names the episode in flight, shows scenes left while re-asking a capped episode, does not re-ask an episode the model merely dropped scenes in, stops the walk on failure while still clearing the scene selection up front and re-reading afterwards, and offers no rewrite once everything is converted.
• client/src/components/fableloom/loomFormats.test.js — the pending-scene/pending-episode helpers.
• Full suites: cd server && npm test (1564 files pass; the one failure, services/loras.test.js > classifies and backfills keyLayout, is an unrelated pre-existing flake that passes in isolation) and cd client && npm test (735 files, 9366 tests, all pass, no act() warnings). cd client && npm run lint clean.

Closes #4794