Four per-render costs: measurements and proposed fixes

Four per-render costs, each with a proposed fix. Three carry a behavior change, one is a ~20-25 file diff, so they are worth agreeing before the patches.

Windows 11, bun 1.3.13, node 26.7.0. Corpus 18,837 transcripts, 16.2 GiB.

1. jsonl-blocks.ts:183 reads past its own 48 h cap. Parses every byte of every selected transcript. A backwards scan stopping at the first pre-48 h entry gives the same answer.

	current	bounded
total	10,443-16,186 ms	2,165-3,056 ms
bytes	1,867 MiB	219 MiB
peak RSS	6.4-8.2 GiB	~460 MiB

Also fixes data loss: readFileSync on a 1,359 MiB transcript throws ERR_STRING_TOO_LONG under node, the bare catch at :221 returns [], 41,652 timestamps vanish (#550's class). Behavior change: with no 5 h gap in 48 h and pre-48 h history in a selected file, the block phase shifts.

2. Git widgets run ten commands where three suffice. 626.3 → 205.3 ms, 21 fewer processes per cold render, via status --porcelain=v2 --branch -z + diff HEAD --shortstat + remote get-url. Spawn is ~90% of a git call on Windows. Three behavior changes: branch.oid at 7 chars ignores core.abbrev; conflicted files are double-counted today (git.ts:433-465); getGitChangeCounts double-counts a line staged then re-modified.

3. chcp.com spawns every Windows render (ccstatusline.ts:325). 35-104 ms, 13-20% of a render; removing it 401 → 288 ms. The code page is console-wide and persistent, so gating on a session_id marker keeps #186 fixed.

4. Widget barrel evaluates React and Ink per render. widget-manifest.ts:5 loads all 88 widget modules; six import ink/react at module scope for renderEditor. 1.57 MB, 55-140 ms, 6-24 MB per repaint. Stubbed to empty modules, output is byte-identical. Fix is ~20-25 files.

Rejected: lazy import() on runTUI (no win, +4 MB, barrel pulls React anyway); directory-mtime pruning (NTFS does not bump on append); batched async stat (getBlockMetrics runs inside a sync render()); fdir withStats (absent on tinyglobby's version).