[mbreiser]

Summary

Two Studio fixes found during the LAB-227 private-repo test on ww18 (2026-10-29, Production 59a377b).

1. Overwrite guard could miss an existing file (every repo). GitHub returns the same ETag for the raw (application/vnd.github.raw) and the JSON form of one contents URL. After a raw read (?p= open, promote, pattern previews via GH.runBytes), Chrome revalidated the cached raw entry for the later JSON GH.run on the same URL, got a 304 and served the raw YAML; res.json() threw, data was null, and resolveRepoSaveTarget (and the silent-fork guard) concluded the file did not exist. A changed protocol then overwrote the repo copy with no "Overwrite?" prompt. Reproduced in the page: the default-cache JSON request returned version: 3…, the same request with cache: 'no-store' returned JSON with sha. Fix: GH.run sends cache: 'no-store'. runBytes keeps caching, since nothing JSON is ever stored now, the collision cannot happen in either order.
2. Telemetry "dropped N" read as in-run loss. The ring block header's dropped is the controller's lifetime eviction count since its boot; a Studio connected hours later logged "dropped 1929" with gaps 0. The drainer now keeps the counter at its first block of a ring incarnation (droppedBase, reset on a fresh ring and by reset()) and exposes droppedSession; the once-a-minute stats line reads "evicted this session N (controller lifetime M)". telemetry_dump summaries and soak stats carry the new field; the raw dropped stays for the record. docs/development/telemetry-logging-reference.md documents it.

Linear: LAB-227 (findings posted there).

Tier

next candidate — together with #285, #286 and #288. The overwrite is recoverable from git history, so not a hotfix.

Release notes

Studio

• Saving over an existing repo protocol always asks first. A browser-cache collision could skip the "Overwrite?" prompt after opening or promoting a protocol; it no longer can.
• The telemetry stats line now separates in-session evictions from the controller's lifetime counter, so a large number after connecting to a long-running controller no longer looks like data loss.

Test plan

• pixi run test — 72/72 files
• tests/test-studio-github.js: run() sends cache: 'no-store'; a non-JSON body yields data: null without throwing
• tests/test-arena-telemetry.js: first block with dropped 1929 → droppedSession 0; next block 1931 → 2; reset() clears the baseline
• On ww18 after the candidate is on Next: open a repo protocol via ?p=, change it, Save → "Overwrite …?" prompt appears

🤖 Generated with Claude Code