feat: confidence-gated Accept/Refine/Restart at the local_triage boundary (Closes #1339)

[Lexus2016]

Closes #1339 — slice B of the AREX loop (#1330), building directly on the StageResult contract merged for slice A (#1338).

The gate

evolution/lib/stage_gate.py — decide() returns a GateDecision:

Branch	Condition	Behaviour
Accept	confidence >= tau	proceed as normal
Refine	below tau, trajectory recoverable	carry evidence forward, re-investigate only the gaps
Restart	below tau, nothing salvageable	discard trajectory, reinitialize from the original problem

tau defaults to 70 — conservative, per the issue's "conservative τ prevents runaway looping" criterion.

There is a deliberate interaction with StageResult.wrap: it assigns confidence 50 when evidence exists but no self-assessment was made. That lands in Refine, not Accept — so a stage that never assessed itself is never silently trusted.

On the recoverable-vs-noisy judgement

The issue specifies this call is "produced by the model running the stage". A deterministic script has no model, so decide() takes an explicit recoverable argument for callers that do have that judgement, and falls back to a structural proxy when it is None: evidence present means something concrete to build on; no evidence is indistinguishable from noise. Both overrides are tested.

Wiring — not dead code

Wired at the same boundary slice A emits from. evolution_local_triage.py runs the tuple through the gate, records the branch in the analysis JSON as stage_gate, and prints the decision line:

[stage-gate] local_triage: RESTART (confidence=50, threshold=70) — below threshold with no salvageable evidence — discarding the trajectory and reinitializing from the original problem

Advisory at this boundary rather than aborting, because local triage is a read-only pre-pass whose output the analysis stage consumes — the branch is surfaced for the next stage to act on. The deterministic dead-code check finds three call sites in evolution_local_triage.py outside the module and its tests.

Success criteria

• Gate wrapper implements all three branches
• At least one stage boundary uses the gate
• Branch decision is logged
• Conservative τ prevents runaway looping

Deferred to slice C (#1340), as the issue specifies: stage_refine_rate / stage_restart_rate metrics and multi-boundary rollout.

50 passing across test_stage_gate.py, test_stage_result.py, test_evolution_analysis_gate.py, test_evolution_local_triage.py — with and without PYTHONPATH; ruff clean.