fix(serenity): reject empty edit text and modernize edit-path tests

[byteclimber]

Summary

Three coupled edit-path (PATCH prompt) fixes in the serenity handlers. Base is feat/serenity-intent-classification, not main.

Fix 1 — Empty/whitespace text accepted on edit but rejected on create

parseUpdatePromptBody did String(body.text) with no trim and only rejected text === undefined, so a PATCH with text: "" or " " passed validation, reached renamePrompt, and got classified/defaulted into a blank prompt. It now trims and rejects empty text with a 400 invalidRequest, mirroring the create path (normalizePromptInput) — including the || '' coercion so a falsy non-string (null, 0, false) is treated as empty too, matching create exactly. parseUpdatePromptBody is shared with handleUpdatePromptSubworkspace, so this one fix covers both edit paths.

Fix 2 — Redundant reclassification on no-op text edits (no behavior change, by design)

handleUpdatePrompt always reclassifies from the new text and rewrites tags, even when the PATCH does not change the text. Decision after review: no behavior change. The handler is not sent the old text, and the upstream provider has no GET-by-id, so it cannot detect a no-op text edit without an API contract change (client sends old text) — deliberately out of scope. Classification must also run before renamePrompt for failure-safety (a classify failure leaves the prompt intact), so renamePrompt's is_updated result cannot gate it (it comes back too late). This PR documents the rationale with a code comment and adds a regression test proving an is_updated: false (no-op) rename still writes tags and publishes — guarding against a future "optimization" that would break the documented contract.

Fix 3 — Modernize stale/broken edit-path tests

The handleUpdatePrompt block had 7 stale tests (not 3) that still stubbed the retired delete+create flow (deletePromptsByIds/createPromptsByIds); they failed at runtime (transport.renamePrompt is not a function) and one referenced an undefined createErr. Modernized them to the current renamePrompt + replace-mode updatePromptTagsByIds flow: in-place edit, drops-falsy/malformed tagIds, 409 collision (throw, no tag write, no publish), non-404 rename error (throw, no tag write), tag-write failure after successful rename (throw, publish never fires, partial-mutation warn logged), and 404 promptNotFound — plus a new negative test that a generic Error with .status=404 does NOT map to promptNotFound. Also added the intent tag to the "recomputes the type tag on edit" assertion. The same stale-test class in prompts-subworkspace.test.js (3 edit-path twins) is fixed too, since this PR is the edit-path fix.

Test evidence

• prompts.test.js: green (was 8 failing → 0).
• prompts-subworkspace.test.js: edit-path tests green (3 fixed).
• Full serenity suite: 1046 passing, 3 failing (was 1031 passing / 14 failing on the base branch before this PR).
• npm run lint and npm run docs:lint clean on the changed files; the changed source (prompts.js) passes type-check.

Pre-existing branch breakage (NOT introduced here, out of scope)

The base branch feat/serenity-intent-classification is already red from the LLMO-6190 dynamic-allocation work, in files this PR does not touch. CI will show these regardless:

• eslint: src/support/serenity/handlers/prompts-subworkspace.js:251 — headroom no-shadow.
• tsc (type-check): src/support/serenity/handlers/markets-subworkspace.js:217 (stale @param headroom) and :537 (Expected 5 arguments, but got 6 — a real call-site arg mismatch to createHeadroomGuard).
• 3 failing tests: dynamic-allocation fronting — retryOnQuota wiring, prompts-subworkspace … dynamic-allocation ON …, and intent-classification.js — observability.

These should be fixed on the feature branch (or by the LLMO-6190 owner) separately.

🤖 Generated with Claude Code

[github-actions]

This PR will trigger a patch release when merged.

[byteclimber]

Folded in the LLMO-6190 headroom base fix (was #2883, now closed) because the two are a mutual dependency — each PR's CI runs the full suite/lint, so neither could go green alone. This branch now also fixes:

• markets-subworkspace.js — generateAndAttachPrompts was missing the headroom param the caller/JSDoc require (TS8024 + TS2554); added it and fronted headroom.ensure({ prompts: texts.size }, …) before the metered write loop.
• prompts-subworkspace.js — ⚠️ behavior change, please scrutinize: removed the redundant pre-publish ensure (the no-shadow build failure + a second getWorkspaceResources read). LLMO-6190 meters once, before the write; the pre-write ensure already reserves the batch. If a second pre-publish top-up was intentional, push back here.
• prompts-subworkspace.test.js — corrected a stale positional call (options → 8th slot).

Verified locally: type-check clean, lint clean, 223 serenity handler tests passing.