Stage strict 0.8 API and release contracts

[jcwal1516]

Summary

• replace brittle source-shape policy with semantic architecture, workflow, release, and ownership checks
• stage the reviewed 0.8 public API transition across the 18 publishable libraries
• make high-level decoding strict by default while keeping lenient recovery explicit and warning only when recovery is actually used
• add the concrete CUDA lossless encoder facade and tighten move-only resident/batch ownership boundaries
• formalize release-manifest tiers, dependency ordering, package-consumer validation, semver baselines, and reviewed breaking-change evidence
• fix CPU and Metal raw Tier-1 writers so a segment ending exactly in 0xFF emits the required stuffed zero byte

Why

The repository had accumulated policy tied to file shape and implementation layout while several user-visible contracts remained implicit. This made maintenance changes disproportionately large and allowed the easiest decode path to select lenient behavior.

The final Metal validation also exposed a separate correctness defect in selective arithmetic coding bypass: a raw Tier-1 segment could end immediately after 0xFF, leaving the mandatory stuffed zero bit absent. Strict decoding correctly rejected that codestream.

Impact

• malformed input is rejected by default through the high-level 0.8 decode APIs
• callers can opt into documented lenient recovery without configuration-only warnings
• reviewed public API and release tiers have explicit compatibility and validation guarantees
• CUDA lossless encoding has a concrete, reportable backend-selection facade
• selective-bypass codestreams produced by CPU and Metal remain strictly decodable when a raw segment lands on the trailing-0xFF boundary

The intentional 0.7-to-0.8 source breaks and the next 0.8.x baseline are recorded in the checked-in API review artifacts.

Validation

• cargo xtask ci
• cargo xtask release-metal --mode full
• cargo xtask package
• cargo xtask stable-api
• cargo xtask semver
• cargo xtask release-integrity
• cargo xtask codec-math-codegen
• cargo xtask public-support --final
• python3 -m unittest discover -s scripts/tests -p 'test_*.py' — 69 passed
• host changed-line coverage: 84.05% (838/997)
• host critical-path coverage: 83.25% (706/848)
• exact CPU and Metal trailing-0xFF regressions

Exact-SHA evidence

Candidate SHA: 340afad9537dfbb91c30780f921c88201464bdf3

• GPU validation run 30403785378
  • CUDA quick validation: passed in 9m1s
  • Metal quick validation: passed in 3m58s
• PR checks run 30403775038
  • all required checks passed
  • hosted Metal compile and pure tests passed in 18m30s