feat(pm): expand MCP, routines, work items, inbox, and skills

[Neonforge98]

Problem

Product-management work in ORG2 was split across disconnected surfaces: Work Items had no custom statuses/properties/saved views, Routines had two competing owners, Team Inbox could not archive or mute, conversation delivery lost failed messages, MCP servers were not wired per provider, and Skills/Quick Actions had no product entry. The batch that closed these gaps had drifted 42 commits behind develop and carried CI-red lints, a flaky lock release, and two data-integrity defects.

Solution

Land the batch as three reviewable commits rebased onto develop (c1d0e2976), resolving conflicts upstream-owner first and fixing every defect found while re-verifying it (details below):

Commit	Scope
feat(pm): expand work item domain and project surfaces	Work Item domain, custom statuses/properties/saved views, Quick Actions, Routine definitions as the single editable owner with pm_routines as a rebuildable execution projection
feat(agent-runtime): expand collaboration tools and execution support	Collaboration tools, execution support, CLI stderr secret redaction, worktree lock liveness
feat(collaboration): unify Team Inbox and conversation delivery	Team Inbox (archive/unarchive, mute categories), conversation delivery with a retained failed row + retry/edit-then-resend

Potential risks

• Ten conflict resolutions in WorkManagement/TeamInbox chose develop's refactored structure and re-wired this PR's capabilities into it; behaviour in those surfaces should be reviewed as merged UI, not as the original diff.
• Archived inbox items have no entry point on develop's UI (see Known gaps) until a product decision lands.
• WorktreeLockGuard now issues an explicit flock(LOCK_UN) in Drop; any caller relying on close-time release semantics keeps working, but the change touches every worktree lock release.
• Cross-org Quick Action id collisions now skip the row instead of failing the pull; a genuinely corrupt sender-stamped foreign-org row still hard-fails.

Rebase notes (develop moved 42 commits)

Conflicts were resolved upstream-owner first: develop's refactored surfaces (WorkManagement split header, TeamInboxListControls with a trailingActions slot, DetailPanelContainer narrow placeholder, EmbeddedWorkItemDetail empty state) are kept, and this PR's capabilities are wired into them rather than re-adding parallel structure. Concretely:

• Mute-category menu ported into develop's TeamInboxListControls action group; this PR's inline filter tabs were not re-added (develop removed in-inbox filter switching).
• headerDispositionAction (archive/unarchive) kept alongside develop's headerAuxiliaryAction (open in browser) — two distinct header actions.
• onWorkItemNameUpdated reporting lives in the DetailPage components; the older EmbeddedWorkItemDetail effect was dropped as this PR's final shape intended.
• ProjectPanelView snapshot regenerated against the merged UI.

Defects fixed during verification

Found by running the gates and rendered E2E against a fresh build (none were caught by the previous pass):

• cargo clippy -D warnings was red on work_run_manifest.rs (type_complexity) — CI-blocking.
• WorktreeLockGuard release was not reliably immediate on macOS: dropping the guard only closed the fd, and under in-process file churn a fresh flock on the same file still returned EWOULDBLOCK (~10–15% of parallel test runs; reproduced by instrumenting the probe, and shown to be independent of the sibling lock tests). Drop now issues an explicit flock(LOCK_UN) before the file closes — 0/40 failures under the same churn. The test also had a cross-process collision (unique_test_dir now includes the pid) and its release assertion now reports "still held" vs "probe error" so a future failure is diagnosable.
• record_org_catalog_rehome added a third bridge row on carrier delete; collab_leave test updated to assert composition, not a magic count.
• Cross-org Quick Action id collision was a hard Err that discarded the whole entity on every pull, forever, with only a warn; receiver-local collisions now skip the row like the sibling property/status/saved-view families already did. Sender-stamped foreign-org rows still hard-fail.
• Routine-created work items lost routine_source provenance after the owner collapse; orgii-routine-changed had no emitter left. Both restored at the producing boundary with regression tests.
• Work-item E2E spec resolved its data home via E2E_ORGII_HOME while wdio publishes ORGII_HOME, so its seed INSERT landed in the developer's real projects.db. Fixed to read ORGII_HOME first.

Known gaps (not addressed here — product calls)

• Archived inbox items have no entry point on develop's UI. This PR adds archive/unarchive actions and the "archived" filter/data path, but develop removed in-inbox filter switching and nothing else sets viewState.filter. Needs a decision on where the archived view is reached.
• autoDetectKey for Claude Code only detects; it does not persist an account, and a saved OAuth account reports health: unknown rather than probing. Pre-existing.

Verification

All five gates run on the final tree of this PR (on develop c1d0e2976):

Gate	Result
pnpm typecheck	0 errors
pnpm lint (oxlint + eslint --max-warnings 0)	clean
vitest --changed c1d0e2976	642 files / 4286 tests passed
cargo clippy --workspace --all-targets -- -D warnings	0 errors
cargo test --workspace --no-fail-fast	88 suites, 0 failed (final tree, after the WorktreeLockGuard fix)

Gate re-run after the WorktreeLockGuard release fix; the lock test that flaked under parallel threads now passes 20/20 (crate, parallel), 40/40 in the discriminating experiment, and in #939's full suite run concurrently with this one.

Pre-existing develop flake, not touched by this PR (git diff develop -- src-tauri/crates/perf-utils is empty): app_memory::tests::macos_region_walk_splits_current_process can fail under memory pressure when two cargo suites run at once (its own comment anticipates page compression); tracked separately.

• Rendered dual-instance E2E (E2E_PROVIDER_MODE=mock): scenarios A and B pass; the remaining failures were triaged as harness staleness on develop (dead selectors such as sidebar-toggle-*, session-comment-toggle-*), a pre-existing product gap (first replay never writes Team Session Blame since 453025a65), and shared-session cascades — none attributable to this PR.