[omartelo]

Summary

Four fixes from a stress run of the native terminal (floods, nvim, Claude Code and Codex sessions, splits, font and layout changes), one commit each.

• A reopened terminal no longer types into its session, and no longer loses output (fix(terminal)). The window wrote the replayed tail into a terminal already wired to the PTY, so the terminal answered the queries in it again (nvim's device attributes, cursor reports) and the answers landed on the session's input: command not found in a shell, stray text in an agent's prompt. Output emitted between the replay snapshot and the terminal's registration was dropped. terminal.Replay now queues the tail on /ws as a replay frame ([0][id length][id][tail]), in line with the output frames and with what the coalescer held folded into it; the window registers the terminal first, takes no output until the replay frame and sends no replies while it writes it. The RPC returns whether a replay frame is coming.
• fontconfig is asked once per face (perf(fonts)). Every terminal renderer built ran four fc-match subprocesses inside the frame. fonts.Locate keeps each answer for the run; the trap this sets (a family installed while lich runs keeps its substitute) is in docs/ceilings.md.
• A font size step resizes the renderers it has (perf(native)). It built a new renderer per open terminal. Renderer gains SetSize, taken in by the next Measure the way a new display scale is.
• Session output is copied once on its way to the window (perf(terminal)). A hidden flood allocated about 7.7x its output in the backend. The coalescer keeps its batch buffer, the replay tail is one buffer grown up to its 2 MB cap and then overwritten, and the outbox copies a batch straight into its /ws frame. Start refuses a session id over 255 bytes, which no frame header can carry, before its PTY spawns.

Measured

A/B against 564f74c89 on a rig (own backend, virtual display, interleaved runs):

	before	after
font size step, 1 / 4 / 8 terminals (p50)	61 / 159 / 290 ms	26 / 28 / 28 ms
first show of a card, to first frame with content (p50)	65 ms	33 ms
replies typed into the session on reopen	3 of 3 reopens	0 of 3
seq 1 2000000, card opened mid-flood	2000 numbers lost, 2 of 2 runs	none lost or repeated

Visible flood throughput and fps, hidden flood backend CPU and idle CPU stayed within noise. BenchmarkHiddenFlood (new) for the last commit: 232 KB allocated per 32 KB read, 2 allocs, down to 34 KB and 0 allocs.

Tests changed

• The respawn_test and sandbox_spawn_linux_test helpers read the tail from the replay buffer: Replay no longer returns it (contract changed).
• TestReplayBufferKeepsOneChunkLargerThanCap is replaced by TestReplayBufferKeepsTheEndOfAWriteLargerThanCap: a write larger than the cap now keeps its last 2 MB (contract changed; a PTY read is 32 KB).
• The transport tests build frames with frameFor, since send takes a frame.

Test plan

• gofmt, go vet ./..., go test ./... at the root and in native/; internal/terminal with -race
• New TestServiceReplayQueuesTheTailBetweenOutputFrames (fails with the coalescer's batch kept on a cut), TestAttachedTerminalStartsFromTheReplay (fails both with output taken before the replay and with replies sent during it)
• New TestLocateAsksFontconfigOncePerFont (fails without the cache), TestFontSizeStepResizesTheRenderersItHas (fails on the rebuild path)
• New TestReplayBufferKeepsTheLastCapBytesAcrossReads, TestStartRefusesAnIDNoFrameCanCarry
• go test ./api -update: terminal.Replay now returns bool