feat(contracts): carry problem detail through edit, graph, diagnostics

[ScriptedAlchemy]

Follow-ups to #2259.

What was wrong

• Source-edit and graph-tool adapters flattened the problem (source-edit and graph-tool adapters drop the typed problem record #2255). tracedecay_mcp::handlers::edit::render_source_edit_outcome and the graph-tool client path turned the daemon's ApplicationProblemRecord into TraceDecayError::ProjectRoute { reason_code, retryable, detail: String }. MCP and tracedecay tool therefore saw a JSON-RPC project-route error: no structuredContent.problem, no legal_actions, and no detail. A missed source-edit writer lock deadline (lock_deadline) reached clients only as a sentence. The graph-tool owner also mapped TraceDecayError::LockDeadline to a generic graph_tool.failed.
• Adoption census test red on master (test(daemon-service): adoption census expects 34 retrieval capabilities, catalog has 41 #2246). The retrieval family composes the graph lookup capabilities that refactor(mcp): answer graph lookup tools through the owner #2251, refactor(mcp): answer files and config through the owner #2273, and refactor(mcp): answer retrieve through the owner #2284 moved onto the owner (affected, ast-grep-search, branch-*, by-qualified-name, changelog, commit-context, config, derives, diff-context, files, find-exact-symbol, grep, pr-context, retrieve, signature). They are real, callable primitive-namespace capabilities, so the production side is right and the test list was stale.
• TypeScript diagnostics facts lived in message text. no tsconfig owns X (searched ...) and The TypeScript producer (...) has not published ... yet were formatted into SafeDiagnostic.message.

What changed

• One renderer. The application-result renderer (render_application_result) moved from the root crate into tracedecay-mcp (application_output::tool_result). A new ApplicationRefusal { operation, binding_id, problem } keeps the daemon's whole problem envelope.
  • Source edit: run_source_edit returns Err(ApplicationRefusal) and render_source_edit_outcome renders it. Every source-edit refusal is now a semantic tool error carrying the record: problem beside the text, which MCP publishes as structuredContent.problem and --json prints. The CLI's mounting resend reads problem.owner_mount_resend_delay().
  • Graph tools: a refusal that carries detail renders whole. A code-only refusal stays the handler's argument or project-route error. Those problems are the owner's own flattening of handler errors (graph_tool_error_problem), with no kind or action of their own. graph_tool_error_problem now maps LockDeadline through ApplicationProblemDetailV1::from_lock_deadline.
• New detail kinds, built only via ApplicationProblem::from_detail:
  • diagnostics_unsupported { file, searched: [{path, present}] }: unsupported, never retried, correct_request. Unsupported problems may now carry detail, in the record decoder, the wire, and the TS SDK.
  • diagnostics_pending { producer, generation }: unavailable, after_delay, retry. generation is the last generation the producer published, or null.
  • Labelled lines for text output: Diagnostics file, Searched tsconfigs, Diagnostics producer, Last published generation.
• Producers migrated: no_tsconfig_problem and the pending arm of unpublished_diagnostics_problem. The producer path is now forward-slashed.
• TS SDK: validates the two new kinds and allows detail on unsupported. It also validates reset_required detail: feat(contracts): carry the reset-required store as typed problem detail #2283 added that kind to Rust but not to the SDK, which refused it as malformed.
• Regenerated dashboard/src/contracts, the schema bundle, and SDK types.ts. contracts:check is green.
• The adoption census list gains the 17 composed capabilities.

Fail before / pass after

• tracedecay-mcp handlers::edit::tests (reverting only render_source_edit_outcome to the flattening): 6 refusal tests fail with a refused source edit renders as a tool result: ... project route error (...). They pass with the fix. The new a_missed_writer_lock_deadline_reaches_mcp_as_its_typed_detail asserts detail == {kind: lock_deadline, resource: "source-edit writer lock", deadline_ms: 30000}, saturated, and ["retry"].
• mcp_suite str_replace_behavior_test::str_replace_behind_a_held_writer_lock_reports_the_lock_deadline_detail holds the real source-edit-transactions-v1/source-edit.lock with an flock, then calls tracedecay_str_replace over MCP. It asserts structuredContent.problem = saturated / application.lock-deadline / after_delay / ["retry"] / detail.kind == "lock_deadline". On master the same call is a JSON-RPC error, so result is null.
• daemon-service adoption_observation::tests::census_counts_each_composed_family_and_omits_uncomposed_families: on master it fails with composed but not listed: [("retrieval", "capability.application.primitive.affected", ...), ...], and passes here.
• Diagnostics:
  • mcp_suite diagnostics_read_test::diagnostics_read_names_a_missing_producer_and_rejects_a_bad_scope pins the full problem record including detail. On master, detail is null.
  • typescript_file_no_tsconfig_owns_names_the_searched_paths asserts {kind: diagnostics_unsupported, file: "scripts/release.ts", searched: [{scripts/tsconfig.json,false},{tsconfig.json,false}]}.
  • typescript_producer_still_running_is_pending_with_typed_detail uses a held fixture compiler (new TypeScriptFixtureCompiler::Held) and asserts {kind: diagnostics_pending, producer: "node_modules/.bin/tsc", generation: null}.
  • CLI --json: core_cli_suite tool_diagnostics_json_carries_the_unowned_scope_detail and tool_diagnostics_json_carries_the_pending_producer_detail.
• TS SDK, run against master's client.ts: surfaces the diagnostics detail an unsupported problem carries and surfaces the reset-required store detail fail (malformed response), and pass here.

Runtime journey (debug CLI from this branch)

Isolated HOME/TRACEDECAY_DATA_DIR, one daemon under systemd-run --user --scope -p MemoryMax=6G -p MemorySwapMax=1G.

$ flock <data>/source-edit-transactions-v1/source-edit.lock sleep 50 &
$ tracedecay tool tracedecay_str_replace --args '{"path":"src/lib.rs","old_str":"12","new_str":"40","dry_run":true}' --json   # 23:50:01 → 23:50:31, exit 1
 "kind": "saturated", "code": "application.lock-deadline", "retry": "after_delay", "legal_actions": ["retry"],
 "message": "The source-edit writer lock stayed busy past its 30000ms admission deadline; retry the operation.",
 "detail": {"deadline_ms": 30000, "kind": "lock_deadline", "resource": "source-edit writer lock"}
$ tracedecay tool diagnostics --args '{"scope":"file","path":"src/lib.rs"}' --json
{"kind": "unsupported", "code": "application.diagnostics.unsupported", "detail": {"kind": "diagnostics_unsupported", "file": "src/lib.rs", "searched": [{"path": "src/tsconfig.json", "present": false}, {"path": "tsconfig.json", "present": false}]}}
$ tracedecay tool diagnostics --args '{"scope":"workspace"}' --json
{"kind": "unsupported", "code": "application.diagnostics.unsupported", "detail": {"kind": "diagnostics_unsupported", "file": null, "searched": []}}
$ tracedecay tool diagnostics --args '{"scope":"file","path":"src/lib.rs"}'
- Problem: `application.diagnostics.unsupported`
- Diagnostics file: src/lib.rs
- Searched tsconfigs: src/tsconfig.json, tsconfig.json
# TypeScript project whose node_modules/.bin/tsc blocks until released:
$ tracedecay tool diagnostics --args '{"scope":"file","path":"src/index.ts"}' --json
 "kind": "unavailable", "code": "application.diagnostics.pending", "retry": "after_delay", "legal_actions": ["retry"],
 "detail": {"kind": "diagnostics_pending", "producer": "node_modules/.bin/tsc", "generation": null}

Local verification

• Lib suites:
  • tracedecay-contracts 420 and contracts_suite 277
  • tracedecay-mcp 391
  • tracedecay-daemon-service 321
  • tracedecay-source-edit 88
  • tracedecay-application: 466 passed, 11 failed. All 11 also fail on master since perf(graph-db): store each code edge as one relation row #2277 (code graph row file-record has the wrong type), filed as test(application): symbol-graph lib tests red after single-row code edges (#2277) #2279. They passed on this change before the rebase over perf(graph-db): store each code edge as one relation row #2277.
  • tracedecay lib mcp::tools::handlers::: 77
• mcp_suite (full, before the perf(graph-db): store each code edge as one relation row #2277 rebase): 576/576. After rebasing over perf(graph-db): store each code edge as one relation row #2277, 572 passed. The 5 failures are read-cost pins that perf(graph-db): store each code edge as one relation row #2277 moved (test(application): symbol-graph lib tests red after single-row code edges (#2277) #2279, comment). After rebasing over feat(contracts): carry the reset-required store as typed problem detail #2283 and refactor(mcp): answer retrieve through the owner #2284, the source-edit, diagnostics, graph-refusal, str_replace, and retrieve subsets passed: 153 + 46 tests.
• core_cli_suite: 144/144. tool_ subset 53/53 after the last rebases. Five tests failed under load or with the host-CLI example not built, and all passed on an isolated rerun.
• TS SDK 32 tests and typecheck. Dashboard: vitest 210 files / 2028 tests, typecheck, contracts:check.
• cargo clippy --all-targets -D warnings is clean on contracts, application, mcp, daemon-service, tracedecay, cli, and sdk, both with and without test-transport/test-helpers. cargo fmt --all -- --check is clean.
• The last rebase over perf(store): compact lexical postings, symbol entries, and graph keys #2287 (store compaction) touched no file in this diff.

Left open

• The parked-code-index refusals the admin sync and background freshness paths raise directly as ProjectRoute are a separate surface, split to fix(mcp): code-index corruption refusals format park facts into a route detail #2280.
• Graph-tool handler refusals stay code-only because the owner flattens handler errors without a kind. Only detailed problems (today lock_deadline) render whole on graph tools.

Fixes #2255
Fixes #2246

[changeset-bot]

⚠️ No Changeset found

Latest commit: 20dd81a

Merging this PR will not cause a version bump for any packages. If these changes should not result in a new version, you're good to go. If these changes should result in a version bump, you need to add a changeset.

Click here to learn what changesets are, and how to add one.

Click here if you're a maintainer who wants to add a changeset to this PR

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-27T03:19:41.064423Z	20dd81a	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 20dd81a8da

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Restore cold-start retries for CLI graph tools

The new .await? propagates code-only graph refusals before this loop can inspect them. execute_graph_tool_surface still returns no-detail problems as an outer Err, and ApplicationProblem::runtime_mounting() has no detail, so a tracedecay tool graph command issued while a cold project is mounting now exits immediately instead of sleeping and retrying as the surrounding comment and previous implementation promise. Handle the outer mounting error in the loop or preserve it as an inner refusal.

AGENTS.md reference: AGENTS.md:L209-L212

Useful? React with 👍 / 👎.