fix(transcript): a past-the-end window resets to the char-budgeted tail; Electron heap report (#1475) - #1492

Part of #1475 (deep dive into PM3's Electron-only renderer heap growth). The full write-up is docs/perf/2026-10-06-electron-heap.md.

Finding: the unbounded growth is not a bridge or IPC leak. It is the stick-to-bottom pin loss that #1491 fixed: a trim's scrollTop clamp reads as the user scrolling up, so the window never trims again.

In a 30-stream Electron soak on main before #1491, the pin was lost at stream 10:

With #1491, two 30-stream runs stayed pinned: heap 39–46 MB, DOM 12–13k and footprint 141–149 MB, flat from stream 6 to 30.

This PR's fix (clampWindowStart): when the store's retention cap drops a long thread's head, an unpinned window whose start was past the new length used to reset to the last 120 entries regardless of size. That mounted 48k nodes at once in the soak. It now resets to tailWindowStart, the character-budgeted tail.

Where PM3's +12 MB came from (all of it bounded):

• the first thread:updated after launch is a full 4 MB transcript, released by stream 6;
• the markdown parse cache reaching its 400k-char budget;
• V8 code-space warm-up.

PM3's "RSS peak 416 MB" was ps RSS noise; the vmmap footprint is the reliable number.

Verified independently, on main merged with this branch

• typecheck exits 0 and vite build passes.
• test:renderer passes 2754 of 2754.
• The new transcriptWindow.test.ts cases fail on main (2 of 8).
• The branch predates fix(thread): a trim's scrollTop clamp no longer unsticks the transcript (#1475) #1491; the merge is clean and keeps it.

Safety: 18 Electron launches on the PM3 zero-leak recipe, each followed by a leak check, with no clone git writes and no recovered/*. The halts were attributed to live activity: an orchestrator worktree, the #1491 push, and a botpoc commit.

🤖 Generated with Claude Code