S2: plan-exit posture switching + the posture indicator, one-click confirm (spec D2)

[aarontrowbridge]

Important

Problem — No posture switching exists (the engine's Tab-switch is the recorded interim); the plan loop ends in a spec with no route into execution.
Approach — Design-of-record: spec-20260907-011500 (rev 2) — D2. Plan completion emits a structured posture_recommendation (mode + reason) derived from the plan's ARTIFACT SHAPE: implementation-shaped → develop; experiment-shaped → research; MIXED-shape → ambiguous — ask (both named, user picks, no silent coin-flip); spec-for-another-agent → the executor's posture when determinable, else ask; abandoned plan → the offer never fires. The posture indicator (fork #297b, now first-class) renders current mode + recommendation with the FULL affordance set: confirm / dismiss / wrong—actually / stay in plan, one click each. plan.auto_switch = confirm | auto (default confirm); auto defers to the TURN BOUNDARY, announces, and preserves state (the plan's draft + pending questions ride along or checkpoint — never dropped at re-bind). Bidirectional: develop/research loop boundaries offer the walk back to plan. Posture changes record on the session (the ADR-0011 vNext contract).

Acceptance Criteria (spec S2)

• The recommendation derives from artifact shape with the mixed/executor/abandoned tie-breaks
• The indicator: current mode + recommendation + confirm/dismiss/correct/stay, one click each
• auto_switch=auto: turn-boundary deferral, announcement, state preservation (tested)
• Mode switch re-binds mid-session per the Tab-switch contract; posture changes land on the session record
• Loop boundaries offer the walk back to plan