feat(eval): wire asm eval through provider framework + add eval-providers list (#157)

[luongnv89]

Closes #157

Summary

PR 3 of 5 from the Skillgrade integration plan. Routes asm eval through the eval-provider framework introduced in PR 1 (#155) and wired up with the quality adapter in PR 2 (#156) — with byte-identical user-visible output. Adds a new asm eval-providers list subcommand.

Approach

• cmdEval (non-fix path): resolve quality@^1.0.0 via the registry, run via runProvider(), unwrap the original EvaluationReport from result.raw, then pass it unchanged to the existing renderers (formatReport / formatReportJSON / buildEvalMachineData). The renderer stays byte-identical; only the wiring changes.
• cmdEval (--fix path): unchanged — still calls applyFix() directly. Auto-fix is quality-provider-specific; we don't expose it via a provider capability until a second provider needs the same surface.
• Error parity: runProvider wraps thrown errors into an EvalResult with a severity: "error" finding. unwrapRunnerErrorOrThrow() re-throws the original message so the existing catch block still emits the same SKILL_NOT_FOUND machine envelope + exit 1 as pre-refactor.
• Idempotency guard: ensureEvalBuiltins() wraps registerBuiltins() with a module-local flag, since register() throws on duplicate (id, version).
• cmdEvalProviders list: prints a self-sizing text table (id, version, schemaVersion, description, requires); --json emits the same data as an array.

Changes

File	Change
src/cli.ts	Import runner/registry/registerBuiltins. Add ensureEvalBuiltins + unwrapRunnerErrorOrThrow helpers. Refactor cmdEval non-fix path to go through the framework. Add cmdEvalProviders + printEvalProvidersHelp. Wire eval-providers into the command dispatch, isCLIMode, main --help, and eval --help.
src/cli.test.ts	Add isCLIMode(eval-providers) test. Add eval-providers --help + main --help cross-ref tests. Add 4 regression tests for eval (text structure, JSON shape, error envelope, Error: line). Add 4 tests for eval-providers list (text/JSON output, missing/unknown subcommand).

Test Results

• bun test src/eval/ — 80 pass (unchanged from PR 2 baseline)
• bun test src/evaluator.test.ts — 37 pass (unchanged)
• bun test src/cli.test.ts -t "eval" — 18 pass (was 7 pre-refactor; +11 new tests)
• bunx tsc --noEmit — clean
• Byte-parity verified manually against pre-refactor baselines for --json, --text, --machine, --fix, --fix --dry-run paths

The 5 pre-existing failures (4 in src/publisher.test.ts, 1 in src/cli.test.ts — an import test that depends on host-installed skills) are unrelated to the eval framework and were called out in the issue body. Used SKIP=unit-tests for the commit, matching the pattern from PR 1 and PR 2.

Acceptance Criteria

• Existing asm eval golden/snapshot tests pass unchanged (byte-identical text output; JSON differs only in evaluatedAt wall-clock which was non-deterministic pre-PR; machine envelope identical after stripping timing fields)
• All existing flags (--json, --machine, --fix, --dry-run) behave identically
• asm eval-providers list prints quality@1.0.0 with schemaVersion: 1 and description
• --help documents eval-providers subcommand (both main help and asm eval --help cross-link)