feat(agent): Evidence-Before-More-Mutation nudge (arm B, env-gated)

[esengine]

Stacked on #7947. First runtime policy of the intervention phase - Arm B of the causal experiment.

Trigger (calibrated, deliberately narrow)

verification_debt_open AND blind_mutations >= 3 AND not_fired_this_turn

Calibration from the two recorded baseline runs: healthy trajectories peak at blind=2 (edit, edit, script-check), so blind>=2 would have nudged 31/49 (63%) of currently-passing runs; blind>=3 touches 4/49 (8%) while still catching the one recorded failure (blind peak 5, nudged at round 10 of 15 with half the runtime left). The slow-reasoning-gap condition is deliberately excluded and reserved for Arm C, keeping B's causal question pure: "does earlier discriminating evidence have value?"

What

• OutcomeTracker counts blind_mutations (mutations since the last discriminating observation; settles with debt).
• applyEBM (internal/agent/ebm.go): stamps ebm_eligible on every arm - baseline carries the same shadow - and, only under REASONIX_EXPERIMENT_EBM=1, appends the nudge once per turn through the guard channel plus an evidence_nudge notice. Nudge copy explicitly permits finishing the minimum coherent multi-file change before checking.
• e2ebench -policy ebm sets the env for suite task runs (arm switch for interleaved A/B).
• Runtime records only the three per-round fields; e2ebench derives the full mechanism chain offline per run: ebm_eligible_round, ebm_fired_round, blind/debt at fire, ebm_rounds_to_check, ebm_ms_to_check, ebm_reasoning_to_check, ebm_check_within_1/2, plus debt_area and blind_peak. The Outcome-shadow report line gains EBM eligible K - fired F (compliance <=2 rounds P%, median rounds-to-check N) - EBM Compliance is the mechanism-health readout that separates "nudge ignored" from "early evidence useless".

Verification

• New tests: blind counter lifecycle (accumulate across rounds, settle on discriminating check), eligibility-without-firing by default, fire-once-per-turn + reset with turn evidence + notice emission, e2ebench chain derivation + render.
• go test across evidence, agent, e2ebench, trajectory, cli, boot, tool/builtin - all pass; gofmt/go vet/repolint clean.
• Note: TestSnapshotterCapturesWorkspaceChanges (pre-existing, untouched by this diff) showed a timing flake under load during verification (300ms poll vs sleep boundaries); alternated branch runs confirmed it fails independently of this change.

Cache-impact: none - the nudge rides tool results in the turn tail exactly like the progress guard; the provider-visible prefix is byte-identical, and with the experiment env unset no prompt bytes change at all
Cache-guard: TestApplyEBMStampsEligibilityWithoutFiringByDefault pins the default-off path; progress-guard tests pin the shared guard channel
Documentation-impact: none - an env-gated experiment policy plus developer-tool report fields; nothing user-facing changes until the experiment graduates, at which point config + docs land together