fix: remediate CLI, provider, and metrics audit findings

[cluic]

Summary

• Restore CLI execution through npm-style POSIX bin symlinks while preserving side-effect-free imports.
• Let an ordinary Providers-page compatibility test select the first Provider through the existing no-start compare-and-set path, without calling readiness-gated activation.
• Make Overview success rates and Provider/model distributions truthful under semantic failures, dropped observations, saturation, overflow, and fixed UTC windows.
• Propagate Provider model policy through Worker snapshots and harden the proxy data plane, Capture, debug output, and anonymous Metrics against streaming, compression, integrity-header, and protected-value failures found by the audit.

Audit findings resolved

Reported defects

1. npm global/POSIX symlinks could make isDirectExecution() compare a real module path with an unresolved symlink path, suppressing runCli().
2. A fresh Providers-page test omitted activateIfNone, so the following explicit activation hit CODEX_NOT_READY while Codex was unconfigured.
3. Overview counted HTTP-level outcomes as successful Responses completions and could lose/conflate model distribution remainder.

Additional findings

• Provider modelMode/modelOverride did not survive every registry -> service -> Worker snapshot boundary.
• Pass-through requests/responses could be fully buffered, cancellation and partial-response settlement were incomplete, and response-start latency began too early.
• JSON/SSE semantic inspection, compressed-body limits, zstd fallback framing, and SSE line/event accounting had correctness and memory-bound gaps.
• Lexical model override could damage unrelated JSON number/duplicate-key/whitespace bytes and retained invalid integrity/signature headers.
• Dynamic Connection tokens could reintroduce hop-by-hop headers.
• Capture truncation/byte totals and debug/Capture redaction did not cover all protected-value placements and reversible percent/JSON/Base64/base64url/hex forms, MIME whitespace, header names, metadata, Metrics model IDs, or undeclared compressed bodies.
• Metrics saturation/dimension eviction could violate persisted and projected conservation invariants; incomplete data could still appear as a precise rate.
• Browser fixtures, legacy model-policy validation, generated UI copy, package-lock root metadata, README release facts, and Changeset coverage had drifted from production contracts.

Verification

• npm test: 519/519 (458 unit-core + 9 isolated Capture + 51 integration + 1 serial core chain)
• Chromium Playwright: 46/46
• Proxy/Capture focus: 43/43
• npm run lint: 33 source files
• npm run typecheck:ui: pass
• npm run verify:ui-build: exact three-file output pass
• npm audit --omit=dev: 0 vulnerabilities
• Package/native/workflow checks: 21/21
• npm pack --dry-run --json --ignore-scripts: exact reviewed 34-file package
• Changesets status against origin/main: minor
• git diff --check, production/staged sensitive-pattern scans, fixed-port cleanup, desktop/mobile visual QA: pass
• Independent final review: no remaining P0-P2
• GitHub Actions: Linux, macOS, Windows, and release preflight pass

Risk and boundaries

Merge risk is L3 because this changes the proxy data plane, Worker runtime snapshots, and Provider lifecycle-adjacent behavior. The PR intentionally remains draft for human review and the repository's external release gates.

Node versions without native zstd conservatively skip semantic inspection for unsupported checksummed or multi-frame fallback cases while preserving byte-exact forwarding. Local tests use temporary roots, synthetic credentials, and loopback upstreams; they do not replace copied-real-history or external-upstream evidence.