Summary

Lifts the idea_lineage operation from local-only to HTTP/OAuth MCP / agent-callable, mirroring find_trajectory's remote posture. Two substantive commits:

Engine federation (feat(engine))

• getBacklinks, getTimeline, searchTakes, searchTakesVector now accept the federated sourceIds[] array in BOTH engines (postgres + pglite), predicate mirroring findTrajectory (array path wins over scalar; neither = no filter).
• searchTakes/searchTakesVector gained real source_id isolation — previously they filtered only by attributed-holder allow-list, which is not a source boundary (cross-source take leak for federated callers).
• getTimeline's 8-case branch collapsed to one composed query.

Op remote lift (feat(operations))

• Dropped localOnly + the runtime ctx.remote reject; one validated sourceScopeOpts scope threaded to all five gather channels.
• Trajectory channel threads remote = ctx.remote === true → world-visibility facts only for remote callers (fixes a hardcoded remote:false that would have exposed private facts).
• Global contradiction trend is omitted for remote callers (can't prove per-finding source scope yet — fail-closed).
• source argument validated against ctx.auth.allowedSources for remote callers (cross-source IDOR closed); local CLI keeps the free override.
• Phase-2 gather uses Promise.allSettled → a failed channel degrades to an empty bucket + partial/errors flags instead of erroring the whole op; wire schema_version 1 → 2.

Scope was set via /plan-ceo-review (Approach A, SELECTIVE EXPANSION) and locked by /plan-eng-review. Target surface is HTTP/OAuth MCP only — NOT added to the subagent allow-list (deferred, see TODOS).

Test Coverage

• New: engine-parity cross-source exclusion tests for getBacklinks/getTimeline/searchTakes on both engines (real Postgres + PGLite); searchTakesVector source-isolation; op-level remote cases (IDOR reject, contradictions-omitted-for-remote, Promise.allSettled partial-on-failure, local-path-unchanged); schema_version 1→2 regression.
• bun run verify: 30/30 checks green (typecheck, jsonb-pattern, key-files-current-state, progress-to-stdout, eval-glossary, test-isolation, all guards).
• All test files touched by this change + the merge resolution: green (195 pass; the one reported failure was a 5s timeout under 12-file concurrency on the pre-existing idea_lineage parity test — confirmed passing in isolation: 15/15).
• NOTE: the full local unit suite (~890 files / 4 shards) was not run to completion in the ship environment (killed at ~20% after 39 min). CI runs the full suite on this PR.

Pre-Landing Review

• CEO Review (/plan-ceo-review): CLEAN — SELECTIVE EXPANSION, Approach A.
• Eng Review (/plan-eng-review): CLEAN — 1 finding (DRY of the federated predicate, resolved: inline + comment, no fragment helper per postgres.js composition limits).
• Outside voice (independent subagent): caught three real gaps the section review missed, all resolved into this PR — searchTakes had no source filter (now fixed), marking find_contradictions localOnly would have broken a shipped op + the skill (reversed — left untouched, root fix deferred), and "remote-callable" ≠ "subagent-callable" (scoped to HTTP MCP).

Design Review

No frontend files changed — design review skipped.

Eval Results

No prompt-related files changed — evals skipped. (operations-descriptions.ts is tool-description text, not an LLM prompt template.)

Plan Completion

Plan: ~/.claude/plans/system-instruction-you-are-working-tingly-moler.md. All build tasks complete (T1–T10 + D-E1). 3 P3 follow-ups deferred to TODOS.md.

TODOS

• Marked complete: Make idea_lineage remote/federated-callable (the P3 this PR delivers).
• Filed (P3): source-scope the cached contradiction trend; subagent allow-list access for idea_lineage; per-remote cost caps.

Test plan

• bun run verify — 30/30 green
• Affected + merge-touched suites green (engine-parity 15/15 in isolation; operations-idea-lineage 47/47)
• Full unit suite — runs in CI on this PR (infeasible to complete in the ship environment)

🤖 Generated with Claude Code