feat(self-eval): baseline persistence + live agent-backed eval runner (Phase A substrate)

[QuintinBotes]

Self-Eval Gate enforcement — substrate (A2 + A3)

The Self-Eval Gate engine shipped in #64 as a tested engine with injected seams. This wires the two production substrates it needs, with zero change to any existing endpoint — purely additive.

A2 — baseline persistence

• SelfEvalBaseline model + migration 0040 — one row per (workspace, suite), NOT NULL workspace_id (tenant-scoped), unique (workspace_id, benchmark_suite_id), storing baseline_rate + resolved/total provenance + a redacted config snapshot. Deferred from 0001_baseline like its FK target benchmark_suite, so the metadata baseline never creates it first (verified on the Postgres stepwise-walk).
• SelfEvalService — workspace_baseline() (the gate's baseline_for lookup), baseline_for_suite(), and record_baseline(overwrite=) with cold-start establish semantics so a regressing run can never lower the bar it defends.

A3 — live agent-backed eval runner

• Lives in apps/worker (the only layer that may depend on both forge_agent and forge_eval, keeping forge_eval agent-free).
• build_coder_tools — a path-safe read/write/list repo tool set.
• agent_solve — checks out a case's base_commit into a scratch worktree, runs a coding AgentRunner (config under test) that only ever sees the public query, and diffs the worktree into the SolveFn file-map the sandboxed runner re-applies before the hidden tests. The agent never sees fail_to_pass/pass_to_pass.
• ProductionEvalRunner — resolves the workspace's private suite, loads its minted cases, and scores via run_self_eval over LocalSandboxProvider. Returns None on any cold-start (no suite / no cases / no BYOK model) so the gate no-ops rather than fabricating a score.

Verification (all offline, no network / no live model)

• SelfEvalService: 6 tests (cold start, upsert, overwrite-guard, workspace isolation, config-snapshot copy).
• ProductionEvalRunner: 8 tests incl. a full end-to-end — a scripted model writes the fix, real git worktrees are checked out from a temp repo, hidden tests run in the local sandbox; a correct config scores 1.0, a still-broken one scores 0.0.
• make typecheck — 585 files, no issues. Migration round-trips clean on the Postgres stepwise walk.

Follow-up PR wires the enforcement itself (A1 config-gate at the AO settings endpoints + A4 POST /runs baseline establishment).

🤖 Generated with Claude Code