Round-6 prototype — offline profile-patch closed loop + Gate-v4 hardening

This release freezes the Round-6 contract implementation (PR-1 → PR-5.1) as a prototype, not a proven self-improving harness.

Included

Gate-v4 admission: admissibility, per-candidate matched controls, deadband, token guard, audit-consistent receipts.
Edit-body sidecar (harness2-edit/v1): content-addressed, atomic, 0o600, GC.
Offline applier (tools/harness2-apply): plan → preflight → human review → atomic write → verify → rollback; refuses desktop profiles.
Gate-B evaluation (tools/harness2-eval, OEP/v1 C/R/I/T): pre-registered holdout, exact McNemar, real-provider-only apply, forged-report rejection.
Meta verified-experience: learns only from real-provider accepted/rejected evaluations with provenance; legacy experience quarantined by default; compose/* is diagnostic only.

npm run verify — 52/52
npm run verify:task 31/31, verify:meta 16/16, verify:loader 5/5
verify:apply 9/9, verify:eval 6/6
npm run r6:verify PASS (13 gate assertions + spike)
pr3:integration / pr4:integration PASS
stage0 v3 archived 20/20 (GATE INVALID — falsification archive, not production evidence)
stage0:r6 negative PASS (0 applied / 20 rejected), stage0:r6:positive PASS (20/20 applied)

No real-provider efficacy evidence is included. pr4:integration is a fake-matrix calibration run. Production apply requires a providerClass:'real' accepted evaluation report.
providerClass is self-declared locally; the tooling cannot cryptographically prove a report came from a real provider.
Real runner token/step metrics are estimates.
Legacy experience is filtered, not migrated; HMR is only marked pending; probe children have no sandbox.
Packages remain private/unpublished; install from source or GitHub spec. No npm publish in this release.