fix(hooks): no-op pact-team-registration write in in-process teammateMode (#962)

[michael-wojcik]

Summary

Gates pact-team-registration so it no-ops the registry write in in-process teammateMode, where all teammates share the lead's session_id and the {session_id → name@team} write collides many-to-one (last-wins), unable to recover which teammate owns a session_id. The mapping is only meaningful in tmux mode (distinct process + session_id per teammate). Closes #962.

What landed

• Guard (pact-plugin/hooks/shared/session_registry.py register()): skip the registry write only on a confirmed session_id == leadSessionId (the runtime structural signal, read inline via a new _read_lead_session_id helper that mirrors task_claim_gate._read_lead_session_id); write normally in tmux and fail-OPEN on every unknown / unreadable / malformed / non-dict / missing-key / unsafe-team topology. Silent, reversible no-op; keys on the structural signal, never a mode flag; zero new shared.* imports (self-contained-leaf invariant preserved).
• Tests: the standing both-modes matrix (in-process sid==leadSessionId → no write; tmux → write + resolvable), 6-branch fail-OPEN coverage, direct _read_lead_session_id helper unit tests + a positive control, a non-mocked pact_context resolve-seam test, and 2 real-CLI subprocess legs. Non-vacuity proven via counter-test-by-revert: a source-only guard removal flips exactly 4 RED cases. Full suite: 9369 passed, 14 skipped, 0 failed, 0 errors.
• Version: PATCH bump 4.4.34 → 4.4.35 (internal hook hardening; no user-visible surface change).

Why route-b (single-file register() guard)

A multi-agent plan-mode investigation + an adversarial feasibility workflow ruled out the alternatives:

• Route-a (gate the directive across the ~20 spawn-prompt sites) is contract-illegal — a static prompt= template literal is rendered by the dispatcher before the teammate process exists, so it cannot read a runtime signal; it could only carry an authoring-time flag, violating the permanent dual-mode "structural-signal-never-a-flag" contract.
• Route-c' (relocate the directive into session_init's SessionStart injection) was investigated and ruled NOT_VIABLE (3/3 adversarial verifiers, high confidence): SessionStart fires once per OS process, not per in-process teammate spawn — in-process teammates fire SubagentStart (→ peer_inject.py), mode-exclusive — so c''s named injection point never runs on an in-process spawn, making per-in-process-teammate gating structurally impossible there.

Route-b (the register() self-guard) is the only viable, contract-legal option, and the discriminator is config-driven (reads leadSessionId from team config.json, never the registry), so the no-op blinds no consumer — all 3 live registry consumers are already fail-safe-to-baseline.

Acceptance criteria

• AC1 in-process no longer performs an un-fulfillable registration (silent no-op) — ✅
• AC2 tmux registration unchanged — ✅
• AC3 all .teammate-registry.jsonl readers enumerated + verified unaffected (config-driven discriminator; fail-safe consumers; non-mocked seam test) — ✅
• AC4 keys on a runtime structural signal, validated under BOTH modes (standing matrix) — ✅
• AC5 no spawn-prompt dispatch site touched (route-b) — ✅ N/A by design