Add per-root Code Map generation controls

[provencher]

Summary

• publish event-driven per-root Code Map status/progress into Agent Mode workspace rows
• add compact pause/resume controls with pending-action and accessibility states
• gate preload, retry, invalidation, setup, and explicit-demand paths while a loaded root is paused

Pause semantics

Pausing is scoped to the current loaded root lifetime. It immediately revokes publication authority, cancels queued/retry/setup/demand work, and prevents new generation admission. Already-admitted engine transactions may drain to the existing safe currentness boundary but cannot publish as current. Resuming waits for cleanup and schedules one fresh preload; unloading or restarting resets the pause.

Validation

• make dev-test FILTER=AgentWorkspaceRootsSidebarStoreTests — passed, 24 tests
• make dev-test FILTER=CodemapPreloadTests — passed, 16 tests
• make dev-test FILTER=CodemapBindingEngineProjectionTests — passed, 16 tests
• make dev-test-list — passed
• python3 Scripts/test_suite_optimizer.py verify-ledger --ledger Scripts/Fixtures/test-suite-contract-ledger.tsv — expected unrelated repository mismatch remains: 5 missing IDs, 0 stale; the three feature rows are present
• contribution commit preflight — passed
• contribution push preflight — passed against the one-commit stacked range
• PR-ready lint — passed (0/1384 files require formatting; SwiftLint strict passed)
• PR-ready full root tests — blocked by existing CodexNativeSessionControllerGoalConfigTests.testAgentModeDefaultCarriesGoalFeatureConfigToStartAndResume goal-default assertions; the exact test reproduces on the untouched stacked base
• make dev-swift-build PRODUCT=RepoPrompt — passed
• live smoke not run because this change was prepared without launching/relaunching the app

Stack note

This PR intentionally targets perf/swift-pipeline-deep-optimization (the retained remote branch for merged PR #608). The hosted comparison is exactly one commit and the 11 codemap feature/test/ledger files, avoiding unrelated history from a direct branch-base comparison workflow. Retarget to main when appropriate after review.

[baron]

Request changes — production design looks solid; two pre-merge items.

1. Confirmed flaky test (hosted shard 3). Tests/RepoPromptTests/WorkspaceContext/CodemapBindingEngineProjectionTests.swift:117 asserts observedStatuses.first == .unavailable(.jobNotScheduled), but projectionSnapshotUpdates intentionally uses .bufferingNewest(1) and the collector Task is unsynchronized. The initial value is buffered at subscription; if scheduleProjectionPreload publishes before the collector’s first next(), newest-1 drops that value and the first observation is .pending(.scheduled) — exactly what shard 3 reported. The contains { .pending } assertion has the same exposure if the collector wakes only after completion. Please make observation deterministic by consuming/acknowledging the initial value before scheduling, or assert only guarantees of a newest-1 stream (eventual .authoritativeComplete and no consecutive duplicates). Keeping newest-1 in production is correct; this is a test-only fix.

2. Retarget to main before merge. The stacked base perf/swift-pipeline-deep-optimization served its review purpose, but merging as-is lands on that retained feature branch rather than main. PR #608’s head is already an ancestor of main, so retargeting should preserve the identical one-commit/11-file content diff. Require green hosted checks on the retargeted merge ref.

A second independent review confirmed the production path is coherent: suspension is installed before detachment, publication authority is revoked synchronously, cleanup is awaited before resume, re-pause cancels an in-flight resume through the transition token, admission is gated across authority/preload/retry/setup/explicit and multi-ticket demand paths, and shutdown/detach/unload clean up the new streams and observer task.

Non-blocking notes only: pause is intentionally in-memory for the loaded-root lifetime; cancelled/superseded idle states can temporarily render MAP …; and rowHasContextLine is now unconditionally true and could be simplified later.