feat(demo): publish exact-bound MockMed v3 presentation evidence

[abrichr]

Outcome

Publishes the first immutable MockMed evidence pack whose presentation media is bound to exact runtime frames and exact source evidence, while keeping raw evidence media unchanged.

This PR deliberately renders product status and target geometry outside the target page. It does not inject overlay state into the application DOM or use presentation pixels as execution evidence.

What changed

• emits a privacy-safe runtime overlay envelope through an out-of-band observation sink
• binds browser target rectangles to exact top-level viewport observations, including frame-scoped targets from the green nested-frame runtime on main
• exports variable-frame-rate WebM clips with exact decoded-frame PTS sidecars
• exports v2 presentation timelines whose events reference exact frame indexes and media hashes
• preserves raw demonstration, replay, and halt media separately from presentation derivatives
• binds each presentation clip to its exact source recording or case/trial/report/outcome/oracle references and SHA-256 hashes
• validates presentation outcome/oracle parity and recomputes every headline denominator from retained case evidence
• publishes mockmed-triage-v3: 6 cases x 3 trials (18 real synthetic-fixture runs)
• retains validator compatibility for immutable v1 and v2 packs
• inventories public video artifacts while keeping every evidence pack and WebM out of the PyPI wheel and sdist

Evidence

• 3 VERIFIED reference runs
• 15 intentional HALTED fault/refusal runs
• 0 silent incorrect successes
• 0 wrong-target actions
• 0 over-halts
• 0 model calls
• runtime network observation: observed in 18/18 reports
• 63 browser-observed navigation/resource entries, all restricted to loopback
• no off-box or third-party egress observed in the isolated campaign
• Tier 1 loopback MockMed system-of-record verification
• screenshots_may_leave_box=false
• exact final 1.75-second hold represented in media PTS and timeline

The pack is first-party synthetic sample data, not customer production evidence. Its manifest records clean source commit 7cc518ee0b83dd571c0902423134a5525635e6b2, generated after green exact-main run 30191346209 at f9091aab0f22b4a65401252b94d648a939da0575.

Validation

• standalone validators: v1 335 files / 18 runs, v2 335 files / 18 runs, v3 344 files / 18 runs
• 46 focused evidence, overlay-runtime, and release-contract tests passed
• Ruff, formatting, and interop-aware mypy passed
• all demonstration/verified/halted PTS sidecars exactly match independently decoded ffprobe timestamps (8 / 20 / 3 frames)
• release-consistency and positive public-artifact inventory gates passed
• exact wheel and sdist built and passed the distribution-license boundary
• archive inspection confirms no evidence-pack path or WebM payload ships in either distribution
• exact-main post-merge CI for sealed-bundle frame compatibility passed
• clean source-archive validation and clean-tree checks passed

Review boundary

This remains a draft for review and CI. It does not deploy or publish the Cloud demo, and this correction does not merge the PR.