feat(evaluation): add resilient SWE-bench Pro console

[frf12]

Which issue or RFC does this PR close?

None (new capability).

Rationale for this change

PowerContext needs a reproducible OFF/ON evaluation to measure the effect of project-scoped context on real coding tasks. This adds a self-contained SWE-bench Pro evaluation console with a durable queue, bounded retries, and evidence-preserving failure handling that runs unattended.

What changes are included in this PR?

• Evaluation package under evaluation/: FastAPI web console and worker over a SQLite-backed durable queue, with a batch/task/attempt model and immutable attempts
• Resilience: task failures never persist a batch pause; bounded automatic retry (5 attempts with a persisted backoff schedule); transient admission gates for usage, filesystem, and dependency health that reopen automatically; lease recovery only after process-lock acquisition, with cooperative cancellation of running work
• Resource lifecycle: per-attempt evidence export (public incident manifest plus a protected private spool) before containers, networks, and workspaces are reclaimed
• Deployment consistency: web and worker publish build revision and schema version, and claiming requires an exact match
• Portable deployment: rendered systemd templates, an environment example limited to documentation addresses, no environment-specific defaults; the optional telemetry integration and evaluation proxy are disabled by default and enabled only through explicit configuration
• Web UI for aggregate reports, task list and detail, and injected-context traces, plus backend and frontend test suites

Are there any user-facing changes?

New optional evaluation/ subsystem only; no changes to core powercontext package APIs. Deployers provide an environment file as described in evaluation/README.md.

How was this change tested?

• Evaluation backend: 1026 passed (non-live), ruff check, ruff format --check, ty check src
• Root project: make check, 480 passed / 7 skipped
• Frontend: 98 tests and a production build
• Cold-start smoke test from the documented environment example; kill/recovery injection and an unattended 24-task stability run on a real deployment

AI usage statement

Codex (GPT-5) implemented this change. Claude Code performed independent multi-round review and verification of the code, history, and deployment behavior.

[Teingi]

LGTM