fix(v2): stop six responses reporting less than the layer beneath them knew - #7097

Summary

The remaining real findings from cubic's review of the v0.8.12 release PR (#7090), all on the v2 API surface. Verified each against the code before fixing; two needed no change.

Six of the seven share a shape worth naming: the layer underneath already knew, and the response threw it away.

• includeFileBase64 silently did nothing without includeOutput. The contract has always said it requires it, and the read honours that — files are projected inside the includeOutput branch alone. Nothing enforced it, so the flag parsed, was accepted, and was then dropped: a 200 carrying no files and no reason why. Now a 400 naming the missing flag, matching how GET /billing/logs refuses a window bound its period will not read. base64MaxBytes rode on the same projection and is covered too.
• POST /files/{fileId}/unzip answered 500 for a bad archive. ArchiveError had no arm in the v2 error policy, though the internal extract route beside it has mapped the same failures to 400/413 all along. statusForArchiveError stays the single place a reason is classified.
• Both workflow-MCP lists truncated silently. The use cases were already returning truncated, and the copilot handler already publishes it — only the v2 presenters dropped it, while nextCursor: null told a reconciling caller the set was complete. The tools list now returns truncated; the servers list returns toolNamesTruncated, kept separate from "there is another page" so it does not fire on every page with a successor.
• A filtered table dispatch was indistinguishable from an unfiltered one. rowIds: undefined documented itself as "every eligible row", but a select-all dispatch has no row list either — and POST on that same path accepts filter and excludeRowIds, so a caller could create a scope the resource then denied having. The compiled filter stays unpublished (it is stored in a different grammar from the predicate the request was written in, and a test had deliberately locked that in); selectAll and excludeRowIds name the distinction without reproducing it.
• GET /files/uploads/{uploadId} always returned file: null, though the contract promises the registered file after finalization and completedFileId was sitting on the record. That endpoint exists for a caller who lost track of a transfer — the one caller who cannot get the id any other way.
• HEAD on a run file 200'd for ids that do not exist. It resolved the file downstream of authorization, so it disagreed with the GET beside it. The lookup moved into context resolution, matching how readTableDispatch resolves its dispatch id.
• v2 bulk table delete audited a display path as resourceId. It projected from the caller-keyed list, so FOLDER_DELETED recorded /Sales where DELETE /api/folders/[id] records the canonical id — two spellings of one resource that no query could join.

Needed no change

• Cross-organization audit-log cursor replay. The scenario needs one caller authorized for two organizations. member carries uniqueIndex('member_user_id_unique') and resolveEnterpriseAuditAccess refuses any organization the caller is not a member of, so it cannot arise. Reasoning posted on the thread.
• Both documentation findings were already fixed on staging by docs: correct two statements that contradict what they describe #7092, after cubic's review ran.

Type of Change

• Bug fix

Testing

• A test per fix, each verified to fail against the unfixed code. Several of these survived because the existing tests could not catch them: the unzip tests all raised OrchestrationError (which every policy already renders) rather than the ArchiveError the use case actually throws, and the HEAD test mocked the use case and made authorize reject — proving only that the route renders a rejection. The new HEAD coverage sits at the use-case level, where the real decision is made.
• Full suites green: apps/sim 33,288 tests, sim-cli 607 tests.
• bun run lint, all 33 audits (check:audits), and type-check in both workspaces pass. generate:openapi rejected the new response fields until their examples were updated, which is the gate working.

Checklist

• Code follows project style guidelines
• Self-reviewed my changes
• Tests added/updated and passing
• No new warnings introduced
• I confirm that I have read and agree to the terms outlined in the Contributor License Agreement (CLA)