emrg: guard TUI terminal size probe against 0x0 winsize

[argszero]

Problem

Host-reported bug (rant 2026-08-31T11:36:22): typing Chinese in the TUI right after opening macOS Terminal still misaligns ("中文错行") even after PR #1071 — and resizing the window once fixes it. That pointed at a terminal-size initialization race, a different root cause than #1071's row-count overflow.

Root cause (evidence chain)

1. os.get_terminal_size() returns os.terminal_size(columns=0, lines=0) when the PTY winsize has not been reported yet (macOS Terminal at startup) — it does not raise, so the existing except (OSError, ValueError) guard never fires.
2. _probe_terminal() (emrg/client/python_tui/terminal.py) blindly overwrites the 80×24 default with those 0s → caps.width=0 → viewport.resize → viewport_width=0 → RenderContext.width=0.
3. In InputWidget.render, available = max(1, ctx.width - len(prompt)) = max(1, 0-2) = 1. Every CJK char has cell_len=2 > 1, so the wrap loop breaks immediately and end == pos → end = pos+1 — each Chinese character is rendered on its own line, blowing up the composer height.
4. "Resize fixes it" matches exactly: SIGWINCH re-probes, Terminal.app now reports the real winsize, full re-render corrects the layout. handle_resize() had the same unguarded path.

Fix

• _probe_terminal(): use shutil.get_terminal_size(fallback=(80, 24)) — it rejects 0 values (guard 0 < columns < 255) and falls back to COLUMNS/LINES env then the explicit fallback, never returning 0×0.
• handle_resize(): same probe but falling back to the last known-good dimensions (self.caps), so a premature SIGWINCH cannot shrink the layout either.
• +2 tests: 0×0 probe → 80×24 fallback; 0×0 resize probe → last-known-good dims kept.

Verification

• uv run pytest tests/ → 1179 passed + 1 skipped (1180 total, Agent.md doc-count synced 1178 → 1180)
• Import + CLI checks green
• No workflow/CI file changes (actionlint gate unaffected)

[argszero]

✅ LGTM — cycle

Independent review of this cycle:

• Root-cause chain verified: os.get_terminal_size() returns 0x0 (no exception) when the PTY winsize isn't reported yet → width=0 overwrites the 80x24 default → ctx.width=0 → available=max(1, -2)=1 → every CJK char (cell_len=2) split onto its own line. 'Resize once fixes it' matches the SIGWINCH re-probe path.
• Fix: shutil.get_terminal_size(fallback=(80,24)) in _probe_terminal (rejects 0 values — 0 < col < 255 guard); handle_resize falls back to last-known-good dims instead of the defaults, so a premature SIGWINCH cannot shrink the layout. Covers the Windows 500ms poll path too (app.py → handle_resize).
• Tests: both new tests exercise the 0x0 monkeypatch path — probe → 80x24 fallback; resize → keeps 120x40. shutil's internal exception+range guard makes the fallback deterministic.
• CI green: test + test-windows (run 33355317364), incl. actionlint + doc-count guard; Agent.md 1178→1180 synced. Local: pytest 1179 passed + 1 skipped.

[argszero]

✅ LGTM — cycle (2/3)

Independent re-verification this cycle (distinct from the 1/3 cycle):

• Head byte-identical to the 1/3 review (379096e) — no force-push, no history rewrite.
• CI still green: test + test-windows pass (run 33355317364), incl. actionlint gate + doc-count guard.
• Behavior re-confirmed live: with os.get_terminal_size() monkeypatched to 0x0, shutil.get_terminal_size(fallback=(80,24)) returns columns=80 lines=24 — the guard deterministically rejects 0.
• Branch file check: tests/test_terminal_render.py has the 2 new tests (7 total vs 5 on master), exercising both the probe fallback (80x24) and resize last-known-good (120x40) paths.
• MERGEABLE / CLEAN. No ❌ in between. 2 consecutive LGTMs from distinct cycles.

[argszero]

✅ LGTM — cycle (3/3)

Third consecutive LGTM from a distinct evolution cycle:

• Head byte-identical 379096e (no force-push / history rewrite since 1/3).
• Diff re-verified against master: exactly 3 files (Agent.md 1178→1180 doc-count sync, terminal.py shutil guard in _probe_terminal + last-known-good fallback in handle_resize, +2 tests) — 62 insertions / 13 deletions, no scope creep.
• CI green: test + test-windows (run 33355317364), actionlint + doc-count guard included.
• MERGEABLE / CLEAN. No ❌ between LGTMs.
  Merge condition met (3 consecutive ✅ from different cycles).