Refactor project ownership, agent context and verification

[Ooooze]

The project previously loaded a large global agent guide and scattered behavior across broad modules, making ownership hard to discover and verification incomplete. This PR completes the eight-stage reorganization: bounded root/local instructions, task-specific documentation and archived history; feature-owned TUI/local-LLM/OS tools; config and filesystem contracts; explicit runtime composition; and turn/step/dispatch/progress owners in the agent loop. Existing public entry points remain compatible.

Verification is now enforced through documentation budgets/link checks, dependency boundaries/cycle detection, complete TSX test coverage with a non-growing debt ledger, and one CI/quarantine registry. A dependency-free process smoke exercises the built CLI and sidecar with actual filesystem operations and SQLite against a scripted local HTTP/SSE backend.

The branch integrates main through a8afe9638. Upstream approval barriers, grounded memory, task recovery/shutdown, prompt preview, trace routing, HTTP keepalive and Windows support are retained in their new owners. Pure coding-mode policy belongs to approval so HTTP does not depend on TUI. Affected test fixtures are completed without suppressions; download-progress tests use deterministic time sampling.

Validation:

• Production lint/build; documentation/import checks and their negative self-tests.
• Full test runner: 1112 suites passed, 13463 tests passed, four existing skips.
• Test types: 2543 roots, 106 TSX tests; existing debt reduced from 848 to 777, no new diagnostics.
• Built CLI/sidecar process smoke: 13/13 scenarios pass, including approval, FIFO, steering, cancel, finish and restart persistence.
• Dependency graph: zero runtime cycles or exceptions.

Roadmap and stage acceptance · fresh integration evidence

The earlier matched-agent experiments are retained as snapshot-specific evidence; this PR makes no general speed or quality claim. Scripted smoke does not establish live-model quality; real Windows execution was not rerun. Research recommendations proposed after completion are outside this PR. .pr-review-56/ and EVIDENCE_ROUTER_MODEL.md are excluded.