fix: reject wrong-typed reasoner envelope fields as malformed output (#4166)

[atomantic]

Summary

processTaskOutput classified a reasoner run as unparseable-response only when the deliverable field failed validation (payload.proposal supplied but unusable). A wrong-typed non-deliverable field — {"analysis": 7}, or a pause with no resolvable target — still passed the envelope check, produced nothing usable, and was recorded as a legitimate no-proposal success. Garbage output was indistinguishable from "the loop correctly had nothing to propose", exactly the hole #2727 closed for proposal.

validateReasonerResponse now reports which envelope keys were supplied but unusable in a new invalidFields array, and the hook keys unparseable-response off that list instead of re-deriving per-field validity itself:

• analysis present and not a string → 'analysis'
• proposal non-null and dropped (bad scope / no title / unnormalizable slug / not a plain object) → 'proposal'
• pause non-null and dropped (no reason, no resolvable target, "this" with nothing to block on) → 'pause'

null/undefined stays absent, not malformed — an explicit proposal: null remains the documented "nothing to propose" answer, and a valid envelope with a valid deliverable is unaffected. Dropping is still lenient and the function still never throws, so the shared contract only gains a field; the previously-returned keys are unchanged.

Caller audit (the issue's explicit ask): the only production caller of validateReasonerResponse is layeredIntelligenceHooks.js#processTaskOutput (plus the layeredIntelligence barrel re-export and the unit tests). Nothing depended on the lenient behavior for a non-deliverable field. Within the hook, a valid proposal overwrites reason on every branch of the filing path, so flagging an invalid pause alongside a good proposal cannot change a filed run's outcome — the new classification only reaches the no-proposal path. The user-facing reason string is unchanged (unparseable-response already has copy in client/src/utils/layeredIntelligenceReasons.js), so no client change is needed.

Also replaces the hand-rolled proposalAttemptedButInvalid check in the hook with the validator's own report (one source of truth for per-field validity).

From the review round:

• The hook logs which fields were unusable (⚠️ Layered Intelligence: <app> reasoner envelope has unusable fields: analysis, pause) — unparseable-response alone can't tell an operator whether the model returned no JSON at all or one wrong-typed key. That reason token feeds retry counting and the hard-exclusion gate, so the diagnostic shouldn't be discarded.
• The client gloss for unparseable-response was inaccurate for the newly-covered cases ("returned no usable JSON" for a payload that is valid JSON) — reworded to cover both.
• A well-formed blockOnIssue: "this" pause dropped only because its proposal failed validation is not double-reported: one root cause, one entry. This sharpens the report only — invalidFields is already non-empty in that case, so the malformed/clean verdict is unchanged.
• The blockOnIssue-asymmetry (a surviving proposal re-derives reason, a surviving pause does not) predates this PR (Base programmatic-I/O task learning outcomes on the output-hook result, not the exit code #2727) and is now stated in the comment rather than silently changed.
• beforeEach now restores the validator mock's default, matching the existing leak guards for gatherSources / computeHardExclusionGate.

Test plan

• New validateReasonerResponse > invalidFields (#4166) unit block in server/services/layeredIntelligence.test.js: wrong-typed analysis; supplied-but-unusable proposal and pause (including "this" with no proposal); absent/explicitly-null fields never flagged; empty for a fully valid envelope; all three collected in envelope order.
• New processTaskOutput cases in server/services/autonomousJobs/layeredIntelligenceHooks.test.js: unparseable-response for {"analysis": 7}, a targetless pause, and a multi-field-invalid envelope; the offending field names reaching the log; and an end-to-end case driving the hook through the real validateReasonerResponse (the suite otherwise stubs it) so the two halves are verified together, not only in isolation. The existing "well-formed envelope that proposes nothing → no-proposal" and "explicit proposal: null → no-proposal" cases stay green.
• Verified every new test fails against the pre-fix source (8 failures) and passes after.
• cd server && npx vitest run services/layeredIntelligence.test.js services/autonomousJobs/ services/layeredIntelligenceBarrel.test.js services/layeredIntelligenceOutcomes.test.js → 6 files, 559 tests passed.
• cd client && npm test → 647 files, 7879 tests passed (covers the reworded reason gloss).
• Full server suite run: the only failures are 55 pre-existing PostgreSQL-dependent suites that fail in this environment because the configured Postgres instance is not running (… require PostgreSQL) — the identical set fails before the change, and none touch Layered Intelligence.

Closes #4166