fix: clear security advisories and preserve failure diagnostics

[jaden3824]

Website dependency checks were blocked by multiple security advisories, discarded session replies lost their reported token usage, and cached GitHub activity could be shown as a fresh live result. This change clears the current dependency audit and preserves those failure diagnostics.

• Patch Next and the necessary Cloudflare dependency chain, refresh the affected transitive lock entries, and pin Miniflare's Undici to its security patch. The resulting clean install reports zero vulnerabilities.
• Record a returned session reply's provider usage before context, binding, and token-ceiling checks can discard it. Preserve the existing terminal rejection and unknown-cost behavior when no valid reply was returned.
• Show the existing unavailable/snapshot state for status: stale activity responses and guard Cloudflare's cache surface for TypeScript and ordinary environments.
• Incorporate the hardened AgentMeasure workflow proposed by roy-tong in ci: run AgentMeasure conformance on the two urusilla vectors #15. Its original full-pin run failed during action extraction because the pinned repository contains an unrelated dangling dependency symlink. Check out only the same commit's action and validator files, retaining read-only permissions, disabled credential persistence, and all five required invariants. Generic AgentMeasure checks continue to complement the existing Urusilla-specific fixture checks.
• Correct the outdated content-bound compiler description and regenerate the bounded decoder QA snapshot and input digests after the README change.

Validation: clean website install and audit (zero vulnerabilities), lint, production build, and TypeScript check passed. All 248 hybrid-runtime tests and 63 contribution ledger/checkpoint/adjudication tests passed. The remaining offline integration batch ran 309 tests with 298 passes, 11 unavailable optional-data/dependency skips, and no failures. Both AgentMeasure fixtures passed their five supported invariants in a credential-free local run at the exact pinned commit. Decoder QA passed 5,232 behavior checks, 101 baseline tests, and 15 QA tests with zero known findings; adoption and decoder artifact digests were verified. The root suite ran 352 tests with no failures and 11 optional-asset skips; all 444 initial-goal tests passed. At final commit 48b62b2, all eight Conformance jobs passed on GitHub, including Python 3.11, Python 3.13, tokenizer accounting, website, cross-runtime, offline integration, strong-codec, and package build (run 36833529473, attempt 2). The separate AgentMeasure conformance job also passed (run 36833534313). The earlier transient dependency-download failure is superseded by this successful execution; canceled duplicate-run records remain visible in the PR check history.

Prepared with Codex agent assistance. These are same-project implementation and regression checks. They add no provider-backed research result, independent reproduction, adoption, or general token-saving evidence. The protocol and frozen language version remain unchanged. The website has not been deployed by this PR.

Refs #15, #16, #17, #18, #19.