feat(bench): harden PowerPlan and final benchmark gate (#18)

## Summary

Separate benchmark power planning from final holdout evaluation and make
both stages deterministic, digest-bound, and fail-closed.

Calibration observations now produce an immutable PowerPlan before the
final schedule is frozen. The final gate accepts only the exact
scheduled holdout identities and suppresses economic statistics whenever
the campaign, plan, schedule, observations, or quality contract is
invalid.

Closes #7.

## Implementation

- Add an offline `experiment power-plan` entry point that:
  - accepts calibration observations only;
  - requires complete `FrontierDirect`/`NeedleMiss` pairs;
- rejects missing, duplicate, mismatched, failed, non-positive,
contaminated, non-beneficial, or zero-variance inputs;
- emits a deterministic `needle.power-plan/2` artifact with canonical
input and artifact digests.

- Bind the PowerPlan to the frozen corpus and campaign commitment, and
bind the final schedule to the exact serialized plan and required pair
count for each route.

- Harden `experiment final-report` so that it:
- loads the frozen campaign, schedule, and PowerPlan using bounded
reads;
  - verifies their raw digests and semantic commitments;
- requires exactly one observation for every scheduled task, route,
split, repetition, pair seed, and arm;
- preserves infrastructure, quality, staleness, worker, discovery, and
recomputation failures instead of filtering them;
- calculates the paired cost ratio and deterministic two-sided 95% BCa
interval only after the complete contract passes.

- Add deterministic synthetic calibration and final-observation fixtures
for both built-in routes, plus adversarial coverage for incomplete
pairs, duplicated identities, stale artifacts, campaign mismatches,
schedule count changes, calibration leakage, invalid costs, and failed
evidence.

- Update the public synthetic campaign, schedule, manifest, PowerPlan,
benchmark documentation, and project status.

This changes the PowerPlan schema from v1 to v2 and the machine-readable
final-gate report to v3. Existing checked-in fixtures were migrated
accordingly.

## Validation

```text
cargo fmt --all -- --check
passed

cargo test --locked -p needle-bench final_gate
8 passed

cargo test --locked -p needle-bench --test publication_surface
3 passed

cargo test --locked -p needle-bench
53 passed

cargo test --locked -p needle-app
84 passed

cargo clippy --locked --workspace --all-targets -- -D warnings
passed with no findings

npm ci
completed; reported 3 high-severity audit findings

npm run build
passed

cargo run --locked -p needle-app -- experiment power-plan benchmarks/corpus/router-cache/synthetic-calibration-observations.jsonl --corpus benchmarks/corpus/router-cache/manifest.json --campaign benchmarks/corpus/router-cache/campaign.json
emitted a deterministic synthetic PowerPlan with no planning failures and 3 required pairs per route

cargo run --locked -p needle-app -- experiment final-report benchmarks/corpus/router-cache/synthetic-final-observations.jsonl --corpus benchmarks/corpus/router-cache/manifest.json --bootstrap-resamples 10000 --seed 42
verified 3/3 pairs per route and failed closed, as expected, because the public PowerPlan is synthetic

cargo test --locked --workspace
one needle-runtime cache-hit test failed during the aggregate run; the same test passed immediately when rerun in isolation
```

No provider calls or paid benchmark observations were executed.

## Risks/limitations

The checked-in PowerPlan and observations are intentionally synthetic
and remain ineligible for an economic claim. This change validates the
benchmark protocol but does not establish real provider-backed savings.

A production claim still requires private calibration and holdout
material, a production-sealed PowerPlan, an approved provider budget,
and a complete provider run.

The aggregate workspace test run observed one isolated cache-hit failure
outside the changed benchmark files. Its focused rerun passed, but the
aggregate run was not fully green.

Bootstrap work and observation inputs are bounded, and statistical work
begins only after the complete frozen contract validates.

## Documentation/evidence

Updated:

- `docs/BENCHMARKING.md`
- `benchmarks/README.md`
- `PROJECT_STATUS.md`
- public router-cache manifest, campaign, schedule, and PowerPlan
- deterministic synthetic calibration and final-observation fixtures

The synthetic fixtures are protocol-validation evidence only and are not
published economic evidence.

## AI assistance and human verification

```text
AI assistance: code, tests, and documentation — implemented the digest-bound PowerPlan and final-gate changes, generated the synthetic fixtures, and exercised the required deterministic and adversarial cases.
Human verification: reviewed the complete 18-file diff, checked artifact and campaign bindings, inspected the fail-closed statistical path, verified the reported commands and outputs, and confirmed that no provider calls, credentials, unrelated edits, or live benchmark artifacts were included.
```

## Ownership checklist

- [x] I read and understand the complete diff and can explain every
material change.
- [x] I personally verified the reported tests and technical or
performance claims.
- [x] I finalized and personally published the commits and this pull
request.
- [x] I checked for unrelated edits, credentials, generated files, and
live artifacts.
- [x] Public documentation and `PROJECT_STATUS.md` are updated when
applicable.

Author: Punisheroot

PROJECT_STATUS.md

@@ -142,10 +142,13 @@ validation.
 - Linux and macOS lack live platform evidence.
 - The accepted economic results are calibrations, not a powered corpus or
   general savings claim. The public v4 benchmark manifest is answer-free and
-  synthetic; the offline protocol and synthetic evaluator fixture are complete,
-  but real private material, a non-synthetic validated PowerPlan, and a
-  provider run are absent. App provider execution remains fail-closed until an
-  isolated executor/broker consumes only ArmLaunch. This does not prove
+  synthetic. Deterministic power planning, exact schedule/observation binding,
+  fail-closed paired analysis, and the two-sided 95% BCa report path are
+  implemented and offline validated. The checked-in plan is reproducible from
+  explicitly synthetic calibration observations and remains permanently
+  claim-ineligible. Real private material, a non-synthetic validated PowerPlan,
+  and a provider run are absent. App provider execution remains fail-closed
+  until an isolated executor/broker consumes only ArmLaunch. This does not prove
   filesystem ACLs or process isolation; callers must keep private bundles
   unmounted and inaccessible to the runner identity.
 - No public beta, support channel, or compatibility window exists.

benchmarks/README.md

@@ -72,6 +72,10 @@ maintainers; no external bundle path is present in public manifests or arm
 launch projections. A real sealed bundle and validated production PowerPlan
 are absent from this checkout, so offline preflight remains fail-closed. The
 synthetic bundle is complete for evaluator tests but cannot enable a provider
-run. A future caller must keep private evaluator material unmounted and
-inaccessible to the runner identity; this protocol does not claim ACL or
-process-isolation proof.
+run. `router-cache/synthetic-calibration-observations.jsonl` deterministically
+reproduces the checked-in synthetic PowerPlan, while
+`router-cache/synthetic-final-observations.jsonl` exercises the exact paired
+holdout contract and its fail-closed synthetic boundary. None of these
+artifacts is economic evidence. A future caller must keep private evaluator
+material unmounted and inaccessible to the runner identity; this protocol does
+not claim ACL or process-isolation proof.

benchmarks/corpus/router-cache/campaign.json

@@ -1,6 +1,7 @@
 {
   "schema": "needle.multi-task-campaign/2",
-  "schedule_digest": "b3:c853b4ed84d8e9119013d788a7c769c63fd77f63a947706fbe4c128b65ed51d5",
+  "schedule_digest": "b3:3ae9b089ba4c7bee7b58ad562a5797872377ef2d986fba30d75651cc567210dd",
+  "task_ids": [],
   "paid_arms": [
     "frontier_direct",
     "needle_miss"
@@ -15,6 +16,7 @@
     "native_subagent",
     "escalation"
   ],
+  "repetitions_per_task": 0,
   "automatic_retries": false,
   "statistical_claim": false,
   "bootstrap_resamples": 10000,

benchmarks/corpus/router-cache/manifest.json

@@ -16,11 +16,11 @@
   "next_pilot_path": "minimal-live-pilot.json",
   "next_pilot_digest": "b3:18a434227eb9abf2381eb0a65348f0dba66301ccba7dbdeb57c7933b078180cb",
   "campaign_path": "campaign.json",
-  "campaign_digest": "b3:082239af1ebe785555801af74c817c5e49c143759cd3e62ef6607f54dadf5e1d",
+  "campaign_digest": "b3:a434abd47f654b222576e60c29f1640fd311d7fa81fa66419018fb55b008329b",
   "schedule_path": "schedule.json",
-  "schedule_digest": "b3:c853b4ed84d8e9119013d788a7c769c63fd77f63a947706fbe4c128b65ed51d5",
+  "schedule_digest": "b3:3ae9b089ba4c7bee7b58ad562a5797872377ef2d986fba30d75651cc567210dd",
   "power_plan_path": "power-plan.json",
-  "power_plan_digest": "b3:8605f040792744907aa94d3ce017a211b774c16521dbe744acb7a1b587647630",
+  "power_plan_digest": "b3:0cfe45b5cca814731619776de15572959720cd576db2c0a8d8cf0bf8d0b561e6",
   "sealed_bundle_schema": "needle.sealed-oracle-index/1",
   "sealed_bundle_digest": "b3:3750e933fa1226fed1892dc4ba3ddad4f84fbf6bd58135d8f26355103dc681c7",
   "tasks": [

benchmarks/corpus/router-cache/power-plan.json

@@ -1,32 +1,33 @@
 {
-  "schema": "needle.power-plan/1",
-  "plan_id": "synthetic-router-cache-v1",
+  "schema": "needle.power-plan/2",
+  "plan_id": "power-plan-1d6bfc9443debec0",
   "manifest_digest": "b3:8be00a443b7e1a57813d8dbe25e6ea22a988f7536bc0f159fee8a58f0fa6cd7a",
-  "campaign_digest": "b3:647300dac4ba3c304b4052a114f1e80b0eb214e278fd4415bbd576c369652864",
-  "calibration_input_digest": "b3:4d27a8427bd3655d59e88c37efadcccfca4a7d82cd4a4324708180aea31289c0",
-  "estimator_revision": "issue-7-structural-v1",
+  "campaign_commitment": "b3:647300dac4ba3c304b4052a114f1e80b0eb214e278fd4415bbd576c369652864",
+  "calibration_input_digest": "b3:1d6bfc9443debec09dab40fdd8b72bd78e6e5e5bf01075def337b31f92c4a104",
+  "estimator_revision": "needle.paired-log-ratio-power/1",
   "alpha_basis_points": 500,
   "target_power_basis_points": 9000,
   "routes": [
     {
       "route": "locate_implementation",
       "baseline_arm": "frontier_direct",
       "treatment_arm": "needle_miss",
-      "pair_key": "task_id:repetition",
-      "observed_log_ratio_mean": -0.4,
-      "observed_log_ratio_stddev": 0.2,
-      "required_pairs": 1
+      "pair_key": "corpus_digest:campaign_commitment:task_id:route:split:repetition:pair_seed",
+      "observed_log_ratio_mean": -0.4337502838523616,
+      "observed_log_ratio_stddev": 0.10900099103037066,
+      "required_pairs": 3
     },
     {
       "route": "trace_state_flow",
       "baseline_arm": "frontier_direct",
       "treatment_arm": "needle_miss",
-      "pair_key": "task_id:repetition",
-      "observed_log_ratio_mean": -0.4,
-      "observed_log_ratio_stddev": 0.2,
-      "required_pairs": 1
+      "pair_key": "corpus_digest:campaign_commitment:task_id:route:split:repetition:pair_seed",
+      "observed_log_ratio_mean": -0.4337502838523616,
+      "observed_log_ratio_stddev": 0.10900099103037066,
+      "required_pairs": 3
     }
   ],
   "validated": true,
-  "synthetic": true
+  "synthetic": true,
+  "artifact_digest": "b3:88cadb0c43e0ac1a39566f5c76fa1a64949b2ab88bdb592aaaafc76b897b22fa"
 }