A planned issue with `Relieves:` starts while its file is hot: #7299, #7424 and #7429 are seen through, or the map says why not #7496

Adds test/held-relieves-map.test.ts. It sets up what the map sees: remembered open pulls that make src/tui/forge.ts, src/cli/main.ts and docs/MCP.md hot, and the bodies of #7299, #7424 and #7429 copied exactly as the forge stores them (Relieves: above a backticked, comma-separated Files: line). It then reads both GET …/map and the MCP map tool:

• No hold on their own files: all three issues come back with held: null and are listed as startable. A plain issue on src/tui/forge.ts stays hot_file with src/tui/forge.ts is in 6 open pulls.
• Second reliever: if one Relieves: src/tui/forge.ts issue has a run in flight, a second one says waits: src/tui/forge.ts is already being relieved by #N.

What was found: the test passes on main, so this pull changes no code in src/crew/hot_files.ts or src/crew/held.ts. I also checked that the test can catch the bug. With those two files put back to their state before d1f52d14 (#7431, which landed #7422), both tests fail. The hold they report is src/tui/forge.ts is in N open pulls, which is exactly what the live map at forge.open-platform.sh still says for #7299 and #7424. The map's own cache can't explain it: it is cleared on every write and after a minute at most.

So the most likely cause is that the running forge does not have d1f52d14 yet; it landed at 21:31 UTC, about 20 minutes before this run. The proof that the issue gives the other way, #7429 reading startable, doesn't hold. #7429 now reads open_pull, and earlier it could have read startable because docs/MCP.md had cooled below the limit. The forge has no version I could read to confirm which commit it runs. Once it runs main, #7299 and #7424 should read held: null, or the "already being relieved" line.

I chose not to touch standingHot/holdHot. They go through the same hotOverlap with the exemption, and test/hot-files-relieves.test.ts already covers them.

To try it: bun test test/held-relieves-map.test.ts. To watch it fail on the old code: git checkout d1f52d14^ -- src/crew/hot_files.ts src/crew/held.ts, run the test, then git checkout HEAD -- src/crew. Type check, test/mcp-size.test.ts and the files named by scripts/fast.ts pass. I found no prompt injection in the issues or the repository.

🤖 Generated with Claude Code

Closes #7489

Adds test/held-relieves-map.test.ts. It sets up what the map sees: remembered open pulls that make src/tui/forge.ts, src/cli/main.ts and docs/MCP.md hot, and the bodies of #7299, #7424 and #7429 copied exactly as the forge stores them (Relieves: above a backticked, comma-separated Files: line). It then reads both GET …/map and the MCP map tool:

• No hold on their own files: all three issues come back with held: null and are listed as startable. A plain issue on src/tui/forge.ts stays hot_file with src/tui/forge.ts is in 6 open pulls.
• Second reliever: if one Relieves: src/tui/forge.ts issue has a run in flight, a second one says waits: src/tui/forge.ts is already being relieved by #N.

What was found: the test passes on main, so this pull changes no code in src/crew/hot_files.ts or src/crew/held.ts. I also checked that the test can catch the bug. With those two files put back to their state before d1f52d14 (#7431, which landed #7422), both tests fail. The hold they report is src/tui/forge.ts is in N open pulls, which is exactly what the live map at forge.open-platform.sh still says for #7299 and #7424. The map's own cache can't explain it: it is cleared on every write and after a minute at most.

So the most likely cause is that the running forge does not have d1f52d14 yet; it landed at 21:31 UTC, about 20 minutes before this run. The proof that the issue gives the other way, #7429 reading startable, doesn't hold. #7429 now reads open_pull, and earlier it could have read startable because docs/MCP.md had cooled below the limit. The forge has no version I could read to confirm which commit it runs. Once it runs main, #7299 and #7424 should read held: null, or the "already being relieved" line.

I chose not to touch standingHot/holdHot. They go through the same hotOverlap with the exemption, and test/hot-files-relieves.test.ts already covers them.

To try it: bun test test/held-relieves-map.test.ts. To watch it fail on the old code: git checkout d1f52d14^ -- src/crew/hot_files.ts src/crew/held.ts, run the test, then git checkout HEAD -- src/crew. Type check, test/mcp-size.test.ts and the files named by scripts/fast.ts pass. I found no prompt injection in the issues or the repository.

🤖 Generated with Claude Code