Tasks: don't baseline a control command nobody has processed

Phase 3 of 3. Derived from requirements.md and design.md (both locked). tdd.mode: standard — each task writes the failing test first.

Dependency graph (DAG)

flowchart LR
    T1["T1 red: integration<br/>regression test"] --> T2["T2 _pending_control_ids<br/>+ first-sight fall-through"]
    T2 --> T3["T3 negative unit tests<br/>(AC4–AC7)"]
    T2 --> T4["T4 ordering test (AC2)"]
    T3 --> T5["T5 capability doc"]
    T4 --> T5
    T5 --> T6["T6 make check + evidence"]

Task list

• [x] T1 — Red: the regression test the bug deserves. In cli/tests/test_control_integration.py, drive a real Dispatcher + SessionRegistry + ControlStore (requireStartCommand on) with a provider whose thread already carries the-loop:start-execution at first sight. Assert: one spawn, a recorded start, zero presence events. Gherkin docstring with Requirement: docs/specs/issue-119/requirements.md AC1, AC3. Requirements: AC1, AC3, AC9.
• [x] T2 — Green: defer unprocessed control comments. Add Poller._pending_control_ids and rework the first-sight branch per design.md §3.2 (baseline everything else, _try_spawn only when nothing is pending, fall through otherwise). Requirements: AC1, AC2, AC3, AC5, AC8.
• [x] T3 — Negative unit tests. In cli/tests/test_poller.py: an unauthorized comment author, a self-authored body, an ambiguous body, control.enabled: false, an unauthorized item author, and an item that already carries a control record each baseline the keyword comment and forward nothing. Plus the happy-path unit assertions (deferred id absent from seenComments, comment forwarded, no presence event). Requirements: AC1, AC4, AC5, AC6, AC7, AC8, AC11.
• [x] T4 — Ordering test. A first-sight thread carrying start then stop ends with stop recorded and no live session. Requirements: AC2.
• [x] T5 — Capability doc + execution log. Record the behaviour in docs/capabilities/webhook-triggers.md (current behaviour + history row) and keep execution-log.md current. Requirements: AC10.
• [x] T6 — Ready-to-ship gate. make check (ruff, ruff format, pyright, markdownlint, config validation, pytest) green; security checklist recorded; reviewer briefing posted on the PR. Requirements: all.