feat(bench): per-round cognition telemetry in trajectory digests

[esengine]

Stacked on #7941. PR 1 of the cognition-efficiency series: pure measurement, no runtime change.

Why

The 49-task tail decomposed into pure model-generation time (0 retries, 0 compactions, ~99% cache), but the trajectory digest could not say what a 29s round bought. Usage events have carried reasoningTokens/completionTokens all along - this joins them per round so slow rounds become attributable and the numbers backfill over every already-recorded trajectory.

What

• Each classified model round now emits a roundDigest: gap ms, tool ms, executor reasoning/completion/prompt tokens, and the round's actions. Executor-only - planner and subagent usage stay in their own buckets.

• Slow-round census: rounds with a gap >= 8s accumulate count/gap/reasoning totals per run.

• delegation becomes its own round classification (task/parallel_tasks/fleet/research). Previously such rounds classified as mutation, hiding an 82s research call inside a bugfix's round ledger. Counted productive, so waste metrics keep their meaning.

• One new report line in suite reports and -mode traj:

  **Cognition** (49 recorded runs): **reasoning** 38,720 tok · **completion** 108,773 tok (806 reasoning/solved) · **output rate** p50 138 · p90 393 tok/s · **slow rounds** (>=8s) 11 = 22% of model time, 3,852 reasoning tok · **delegation** 2 rounds (1m23s in subagents)

  (real backfilled output over the recorded 49-task run)

• intervals.go extracted from trajectory.go (interval math, unchanged) to stay under the file-size ceiling.

Verification

• go test ./cmd/e2ebench/ (new fixture test: per-round token join, subagent exclusion, delegation classification, slow census, render line) - all pass
• gofmt, go vet, repolint clean
• Backfill exercised against the full recorded 49-task trajectory directory

Documentation-impact: none - e2ebench is a developer benchmarking tool; the new report line and JSON fields are self-describing and not covered by docs/*.md

[esengine]

Self-review pass over the stack found one real defect here, now fixed in the amended commit:

Slow-round census counted recovery/compaction-tainted gaps. A 30s provider-retry gap would have entered the census as a "slow thinking round" carrying near-zero reasoning tokens, misattributing adapter flakiness as cognition. The census now requires the gap to be untainted (!gap.tainted && !gap.compaction); the round's gap itself stays booked in the round digest and outcome tallies, so nothing is hidden — it just no longer counts as a cognition purchase. Covered by TestSlowRoundCensusExcludesRecoveryGaps.

Also verified during review: executor-only token attribution (planner/subagent/sidecar usage excluded before accumulation), digest/actions ordering via the batch's dispatch-order names slice, and the delegation round class staying inside productiveOutcomes so existing waste metrics keep their meaning.