fix: preserve private source metadata and close codec boundaries

[iperev]

Summary

• Preserve direct source/member non-claims, external rule references and proof-binding references in the private source-v2 candidate model and codec, with exclusive profile/member ownership.
• Close UTF-8 path, full-ID diagnostic, composed-text, expanded-budget and absent-payload boundaries without changing public source-v1 or CLI contracts.
• Bind minimum JSON nesting to complete canonical output; route named non-claim collisions to their actual lexical reference owner.
• Add exact projection, whole-operation, lexical-span, resource-precedence and negative round-trip tests. Prepare the compatible 0.14.4 release record.

Validation

• go test -mod=readonly ./... -count=1 passed.
• Targeted model/codec/release-change tests and model/codec go test -race passed.
• Five retained isolated mutants failed at the expected oracles: three absent-field guards, the nesting floor and the source diagnostic operand.
• Independent architecture, authority, decomposition, governance, checklist and test-oracle reviews were followed by bounded repair reviews. The final source was checked unchanged against the frozen 97-file review manifest before commit.
• Full committed npm run check passed on unchanged 82b3e1362760fc1b533a0e4b6bab839df7ada2d0 with a clean worktree, including 315/315 Chromium, Firefox and WebKit tests, package artifacts, self-hosting receipts, coverage and release closeout.
• Exact-head provider checks remain a separate mandatory gate; merge waits for every applicable check and release waits for postmerge verification.

Scope And Limitations

The candidate remains private. This PR does not perform public source-v2 cutover, change supported platforms, establish a DSL/token-cost advantage or prove consumer adoption. Generated review results do not prove host-enforced agent isolation or an exhaustive absence of defects; those stronger campaign claims remain unverified. The previously missing instrumentation-byte review is closed by two independent reviewers, separately from native execution receipts.

Retro

Earlier review found a minimum-depth inconsistency, a wrong lexical diagnostic owner and a rejection-only test that did not observe pre-rejection copying. Their original failures were retained and the repairs were tested causally. A narrow follow-up supplied complete instrumentation variants and overlay mappings instead of relying only on logs. No installed skill was changed.

The first committed full gate stopped because the new local worktree lacked its locked TypeScript development dependency. Installing the unchanged lockfile with the declared npm version and scripts disabled restored the prerequisite; the complete gate then passed without a source change. The failed attempt remains retained separately from the passing successor.