Add osaurus bench --gauntlet: measured model-capability probes writing the ledger

[mimeding]

What

Adds osaurus bench --gauntlet, a seven-probe model verification workflow:

1. Non-empty load/stream response
2. Template-marker leakage across content and reasoning channels
3. Forced tool call with exact function name and schema-valid arguments
4. Multi-character stop handling with an unstopped control run
5. Long-generation degeneration detection, including two-token loops and character runs
6. MTP-versus-AR comparison evidence
7. Prefill/decode performance snapshot

Results are field-merged into the capability ledger without deleting existing verdicts, block reasons, digests, or unknown fields.

Safety contract

The gauntlet records per-probe evidence and a production-serving candidate result. It does not change productionServing in v1 because the runtime cannot yet bind that promotion to a verified weights digest. Failed or unproven critical probes exit nonzero. MTP equivalence remains untested unless the runtime can prove per-request MTP activation.

Validation

• BenchGauntletTests: 20 passed
• DegenerationDetectorTests: 12 passed
• Scoped SwiftLint: passed
• git diff --check: passed
• Independent runtime review findings addressed: no vacuous stop/template passes, no stale ledger promotion, no destructive record replacement, and no two-token degeneration blind spot

Risk and rollback

CLI and evidence-writing scope only. Runtime model behavior, prompts, templates, samplers, and generation defaults are unchanged. Reverting the single feature commit removes the workflow.

[mimeding]

Adversarial-review hardening pushed (e275312): stop-sequence probe now requires positive evidence (counted to the boundary) and fails on leaked stop text; mtp-equivalence and the canary guard against vacuous empty-transcript passes via real usage token counts; --tune-prefill/--gauntlet mutually exclusive. 15 new tests.