fix: cap aggregate untracked content in review context (#405)

[ayobamiseun]

Fixes #405

Problem

collectReviewContext bounds the inline diff (256KB) and each untracked file (24KB), but never bounds the total Untracked Files section. The section is also inlined even in the lightweight includeDiff: false path, so the maxInlineDiffBytes heuristic never protects it. With enough untracked files the two independently-bounded sections compose into an unbounded prompt and adversarial-review dies with the opaque model-layer error:

Input exceeds the maximum length of 1048576 characters.

Repro

Repo with 300 untracked text files just under the 24KB per-file cap (mirrors the issue's real case of 309 files):

	assembled context
before	7,086,684 chars (6.7x over the 1,048,576 limit → exit 1)
after	271,243 chars (review proceeds)

Fix

New formatUntrackedSection tracks a running byte budget across the untracked entries — default 256KB (DEFAULT_UNTRACKED_TOTAL_MAX_BYTES, matching the existing inline-diff cap), overridable via collectReviewContext(..., { maxUntrackedTotalBytes }). Once the next entry would exceed the budget, it stops inlining and appends a visible notice instead of failing downstream:

(N untracked file(s) omitted: aggregate untracked content exceeds the 262144 byte limit. See the Git Status section for the full file list; inspect the omitted files directly, or commit or gitignore them to narrow the review.)

Nothing becomes silently invisible: the Git Status section (--untracked-files=all) still names every untracked file, and skip markers (directory/binary/oversized) count toward the budget too. This is option 1 of the fixes suggested in the issue.

Tests

• collectReviewContext caps aggregate untracked content instead of growing unbounded — cap kicks in, omission notice present, Git Status still lists omitted files.
• collectReviewContext keeps all untracked content when under the aggregate cap — no behavior change for small trees.
• Full suite: 92/93 pass; the one failure (resolveStateDir uses a temp-backed per-workspace directory) fails identically on main when run inside a live Claude Code session (pre-existing Test suite is not hermetic inside a live Claude Code session: 4 failures at HEAD + fixture state leaks into the user's real plugin data dir #455, unrelated).