Add release-tag measurements and capture recovery evidence

[jakezwang]

The previous benchmark evidence did not measure a published v2 Go module or sustained capture across an API crash. This change records a clean, frozen v2.1.2 release-source matrix and a separate checksum-verified release-binary recovery experiment, with raw samples, source manifests, environment and explicit measurement limits.

• All 12 cells passed: 1k/10k/50k documents × 1/4 workers × 1/4 ancestry levels, including 1,200 metadata forks, 240 sandbox/capture workflows and 15,000 verified captured divergence updates. Reported quantiles were independently recomputed from raw samples.
• The 600.25-second single-writer recovery experiment verified 1,865 acknowledged native writes across 1,866 barriers, including a write while the API was stopped. Restart-to-verified recovery was 0.329 seconds; final graceful shutdown was 0.039 seconds, exit 0. This is bounded local correctness evidence, not production tail latency or database failover evidence.
• The runner uses valid /v2 modules, exact release refs, frozen source provenance and explicit quiesced imports. The new recovery harness retains failure evidence through cleanup errors and removes only its own disposable databases. CI runs the current release in an isolated replica-set workflow smoke.

Validation: Go tests/vet, real replica-set smoke, full recorded release matrix, independent raw-sample checks, real recovery experiment, and failure-path/cleanup checks. Historical results remain labelled; the 1m-document/16-worker/16-level matrix is still unmeasured.