fix(serenity): finish LLMO-6190 headroom wiring (unbreak type-check + build)

[byteclimber]

Summary

The feat/serenity-intent-classification branch shipped the LLMO-6190 dynamic-allocation headroom guard half-wired, which is why the branch's own CI (PR #2785) and the stacked edit-path PR #2875 are both red. This fixes the base breakage. Base is feat/serenity-intent-classification, not main.

What was broken (reproduced on the pristine base)

• type-check: markets-subworkspace.js — generateAndAttachPrompts's JSDoc documents a required headroom param and the caller passes it, but the function never declared/used it (TS8024 + TS2554: 6 args to a 5-param fn).
• ci / build (eslint): prompts-subworkspace.js:251 — a second headroom guard shadowed the one at line 171 (no-shadow).
• test: handleCreatePromptsSubworkspace dynamic-allocation test failed — its options object was passed in the env positional slot instead of the 8th, so the guard was never enabled.

Fixes

1. markets-subworkspace.js — add headroom as the 6th param to generateAndAttachPrompts and front headroom.ensure({ prompts: texts.size }, { includeDrafted: true }) before the metered createPromptsByIds write loop (the metered write is the create, not publish). Matches the JSDoc + the prompts-subworkspace.js twin. No-op when the flag is OFF.
2. prompts-subworkspace.js — ⚠️ behavior change, please scrutinize: removed the redundant pre-publish ensure (which re-created a second guard → the no-shadow, and issued a second getWorkspaceResources read). The LLMO-6190 design meters once before the write (test title: "front headroom before the write, not just before publish"; assertion: getWorkspaceResources called once, before createPromptsByIds). The pre-write ensure at line 176 already reserves the whole batch, so the pre-publish read was a pre-LLMO-6190 leftover. If a second pre-publish top-up is intentional, this is the line to push back on.
3. prompts-subworkspace.test.js — fixed the stale positional call (options → 8th slot) to match the controller + signature.

Verification

Type-check clean, lint clean. Combined with PR #2875, the three serenity handler test suites are 223 passing / 0 failing.

⚠️ Interdependency with #2875

This PR and #2875 are complementary — the feature branch is broken in two places and each PR fixes one, so neither is green alone; both must merge for feat/serenity-intent-classification (and PR #2785) to go green. Verified locally that edit-path + headroom together are fully green.

🤖 Generated with Claude Code

[byteclimber]

Consolidated into #2875. These two fixes are a mutual dependency (each PR's CI ran the full suite and was red on the other's unfixed domain), so two separate green PRs was technically impossible. The headroom/LLMO-6190 commit is now cherry-picked onto #2875; that branch is green (type-check + lint clean, 223 serenity handler tests passing). Closing this one.