docs: publish consolidated cache, startup and CI cost evidence

[zackees]

Summary

Refs #1286. Publish the consolidated follow-up evidence without claiming the unfinished audit complete:

• Real Tier 2 cold/warm reseed observations and same-run artifact transfer costs.
• All eight expanded setup-soldr lane setup/post observations, preserving separate native ccache and embedded soldr counters.
• Synthetic pinned janitor executor evidence: global-budget reduction alone does not implement global eviction; fixture limits explicitly recorded.
• Observed canceled/failed runner time for the two recovery/startup PR branches, with snapshot limits and active runs excluded.
• Successful startup PR and merge-queue verification: 295 tests per mode, no failures; local startup reproduction remains separate.
• Linux release host namespace landing ci: align Linux release cook cache with host namespace #1367, without claiming an unperformed release benchmark.
• Latest ci: preserve cache failure evidence and compile through daemon outages #1362 validation: build succeeded; both modes failed unchanged recording-shutdown state assertions, tracked under existing frontend: RecordingShutdown polls every 10 ms on the UI thread while waiting for recording to stop #1283. No blind reruns or assertion changes.

The report retains unresolved full-nightly warm Ubuntu reuse, representative-day traffic, global cache pressure, deep restored-target freshness, and daemon root cause. Cache publication architecture remains a proposal. #1368 separately extends the known unsafe Rust-wrapper bypass.

Validation

Documentation/evidence only. Every changed JSON file parses; git diff --check passes. No new workflow, source change, or submodule pointer update. No artificial RED/GREEN test required for recorded measurements.