feat(evidence): shadow outcome scorer with trajectory recording and bench metrics

[esengine]

PR 1 of the outcome-progress series: pure measurement, no behavior change.

Why

The adaptive progress guard (#7915) scores novelty - new reads, new commands, mutations. That conflates motion with progress: an agent editing a different file each round earns +3 forever while getting no closer to a passing check, and a delegation return earns +2 whether or not it produced usable evidence. Before changing the policy, this PR makes the gap measurable.

What

• evidence.OutcomeTracker: a shadow scorer observing the same per-round receipts as the live ProgressTracker, decomposed by outcome instead of novelty:
  • exploration - first reads, first commands/actions, first failures (localization), delegation returns
  • verification - attempts of verification-classified commands (IsDeliveryVerificationCommand, the same single classifier readiness uses)
  • objective - a verification command turning fail->pass
  • regression - pass->fail
  • churn - successful mutations (a state transition, not proof of progress)
  • legacy_gain - the live scorer's verdict on the same receipts, so offline analysis can price its claims without replaying
• New optional sink capability RecordOutcomeProgress (mirrors the contract-shadow pattern from feat(taskcontract): unified task record assembled from existing signals #7927); the trajectory recorder appends each sample as an outcome_progress record. Pass-through added to every capability-forwarding sink decorator.
• Fixed pre-existing pass-through gaps found while wiring: notify.Sink dropped readiness audits, turn completions and contract-shadow audits; the telemetry sink dropped contract-shadow audits. With notifications or telemetry enabled, those records never reached the trajectory recorder.
• e2ebench condenses the series per run into an outcome digest: false_progress_rounds (legacy gain never redeemed by an objective transition within 3 rounds), solution_stall_max (armed only once the run enters its solution phase, so research turns are not punished), best-vs-final verified state, regressed_from_best, and search_regret_ms. Suite reports and -mode traj gain one Outcome shadow aggregate line.
• Recordings that predate the shadow scorer backfill from shell verification receipts (execution.verification on tool results), so existing trajectory corpora can already answer how many runs peaked above their final verified state - the number that decides how much PR 2 (best-known-state) and PR 3 (conservative rollback) are worth.

Verification

• go test ./internal/evidence/ ./internal/event/ ./internal/trajectory/ ./internal/agent/ ./internal/notify/ ./internal/telemetry/ ./internal/stats/ ./internal/control/ ./internal/cli/ ./cmd/e2ebench/ ./internal/tool/builtin/ ./internal/boot/ - all pass
• TestOutcomeShadowRecordsEveryRoundWithoutTouchingGuards proves guard-visible round outputs stay byte-identical with the shadow active
• gofmt, go vet, repolint clean

Cache-impact: none - the provider-visible prefix and all injected guard texts are byte-identical; the agent change is a private struct field plus a sink capability call after each batch
Cache-guard: TestOutcomeShadowRecordsEveryRoundWithoutTouchingGuards plus the existing progress-guard tests pin the guard's injected texts; no prompt bytes are produced by the shadow path
Documentation-impact: none - shadow measurement is invisible in the product; the trajectory file gains an optional record type and e2ebench (a developer tool) one report line, neither covered by docs/*.md