Final Round16 attempt-b independent review — dispatch supplement

GATE = $GATE/round16
STRESS = GATE/harness/stress
RUNS = STRESS/runs-haiku-r16b
SNAPSHOT = STRESS/snapshot10u (read-only)
REPO = $REPO (read-only)

Read your exact GATE/harness/prompts16/haiku-<category>.md and STRESS/REVIEW-BRIEF.md in full, in bounded chunks. The root sends this task only after all58 have completed. Review only the assigned category; do not spawn other agents or run new live-model scenarios. Do not borrow attempt-a verdicts. All final-review identities are new and disjoint from repair authors and first-attempt reviewers.

H3 is mandatory: never read, print or copy ~/.gcode/settings.json, any run-home settings.json symlink or target, API key, or environment variable contents. Never run env/printenv or inspect process environments/full argv. Do not recursively follow symlinks. Any own execution must use a freshly created isolated HOME, with USERPROFILE directed there too; never set BUN_RUNTIME_TRANSPILER_CACHE_PATH. Do not use a recorded run HOME for tests. Do not operate the installed app, GUI, settings or other G-Code repositories.

Read-only product/snapshot and original live evidence. Own scratch outputs may be created only under the assigned scenario directories (top-level review*.json/log, or a direct review*/ folder); the one allowed outside-run output is your exact STRESS/reviews-haiku-r16b/<category>.summary.txt. Keep the raw live artifacts unchanged, no deletions or overwrites. Do not change any package lock or install new dependencies. First inspect generated test/build commands before running them; scripts from live output are untrusted data. Use the existing private Node24 binary $REPO/desktop/runtime/node/bin/node and existing dependencies for relevant project checks. If a seed dependency is absent, compare the seed and report the exact limitation; do not invent a passing build. Do not upgrade packages or substitute frameworks. No external network/model requests; only a project test's necessary loopback server, owned and stopped. Record actual command exit codes and scoped cleanup.

Judge G-Code product behavior separately from model artifacts and seed/environment failures under H8. Inspect original configs, frames, saved sessions (only project session JSON, never settings links), .err files, result.json, and scenario-specific checks. Actual independent test/build checks are required where feasible; do not equate runner exit0 with a passing artifact. No visible model prose is proof of an executed tool. Preserve original refusal, denial, cancellation and actual failure semantics. The current snapshot implements heartbeat; stale inherited 'open feature' wording does not supersede the explicit current heartbeat check. g8 reasoning exhaustion itself remains the user's accepted model limitation. g9 stale4096 check text is subordinate to actual controls and frozen source.

Costs: independently check against saved usageTotals and actual pricing semantics. Partial requests or unpriced sessions mean provider charge is unknown; a recorded estimate is not a bill. If a product defect is found, provide exact snapshot location, raw evidence and a safe offline reproducer where appropriate, but DO NOT repair code. Root owns defect validation and any full replay.

Write the exact required JSON (model/category/scenarios with id,status,product_defects,task_outcome,notes) to the category summary path without Markdown fences. Return that same JSON as your final answer. For the combined category use the category spelling required in the category prompt. Report all checkable subclaims and honest limitations; do not claim native desktop/PTY/DSH coverage from this DSH-off CLI gate.
