fix(scenario): route the evaluator's report emits through write-report.js

[pjpoulose]

Routes scripts/scenario.js's report emits through scripts/write-report.js, closing the
last instance of the report-loss class fixed across every other CLI in 0da6f20.

Why this is a separate PR

scripts/scenario.js is the evaluator file. The separation guard (EVALUATOR_EXACT) rejects
a PR that mixes it with behaviour changes, so this cannot ride along with the cleanup branch.
Guard verdict on this diff: pass, 1 file, no mixing.

What it fixes

One helper, out(), backs all three emit sites — out(scenario) in record, out(bundle) in
replay, out(result) in selftest — so a single line closes them.

The path that can actually lose bytes today is selftest: out(result) is followed immediately
by process.exit(1) when the run has errors, and process.exit() discards whatever libuv
still had queued. The exit code survives, so this never fabricated a pass — but the report
explaining why the evaluator failed is exactly what got truncated.

The replay and record paths return instead of exiting, so Node drains them. They are routed
anyway, because the guarantee belongs to the emitter rather than to the caller's control flow,
and a later refactor that adds an exit must not silently reintroduce the defect.

Honest scope of the measurement

Measured on Linux / node 22: file 673B, spawnSync pipe 673B — identical. No loss is
reproducible on that platform at this size, and that is stated rather than dressed up: the
report is far below the 64KB Linux pipe buffer.

The change still matters. The defect is not size-bounded — it produced a completely empty
report in CI, meaning libuv's first try_write moved zero bytes — and macOS's effective pipe
buffer is roughly 8KB. Preventative on Linux, load-bearing elsewhere.

A correction to an earlier write-up

An earlier note flagged line 216 as a second emit site. That was wrong. It sits inside
generateManager()'s template string — a log line in a generated fixture manager, not a
scenario.js report emit. scenario.js derives verdicts from on-disk state, never from output
strings, so losing those lines cannot fabricate a verdict. It is also a generated standalone
file, so a require there would hit the same MODULE_NOT_FOUND trap that keeps watchdog.js
inlined. Deliberately left alone.

scaffold.js copies only watchdog.js into generated projects, so scenario.js can safely
require the helper.

Verification

• scenario.js --selftest exit 0; gate.js --selftest exit 0 (gate consumes scenario output)
• tests/scenario/run-tests.js green
• tests/report-integrity 22 PASS / 0 FAIL, with its positive control firing (110,615 of
  402,967 bytes lost through spawnSync), so the suite's green result is meaningful on this
  platform rather than inert
• graphlint --selftest, graphlint scripts/, docs-lint all exit 0
• CI run #80 on this exact commit: success on all legs (ubuntu 18/22, windows 18/22, macos 18/22)

Not included

The timing-fragility work tracked in claude/graphsmith-ci-timing-fragility-decision.md is
unrelated to this change and lives on the other branch. This PR contains no timing-dependent
code.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.