Refactor: 17% LOC reduction

[tmck-code]

Context

LOC reduction across claude/ (yas) and test/, without deleting working code or tests.

	before	after	change
claude/ (yas)	14,160	11,749	-17.0%
test/	22,073	20,955	-5.1%
combined	36,233	32,704	-9.7%

Changes

Where the yas reduction came from

lines	share	what
-2,231	15.8%	comments and docstrings: removed narration of past changes, deleted docstrings the function name already covers, compressed invariant notes to one line
-577	4.1%	dead code (Pill.gradient_fg, rainbow_color, _parse_iso_to_epoch), spark_rgb/gradient_rgb dedupe
-15	0.1%	parsecache eviction fix
+412	-2.9%	reverting @dataclass to __slots__ classes (see Benchmark)

Where the test reduction came from

lines	share	what
-655	3.0%	151 banner comment blocks, 282 bare # setup/# run/# expected/# assert markers, and their blank lines
-289	1.3%	parametrising and table-driving same-shaped tests
-256	1.2%	_silence_dynamic lifted to conftest (3 drifted copies), hoisted strip_ansi imports, deduped Renderer() instantiations
-31	0.1%	two byte-identical duplicate tests
+113	-0.5%	fixing three defective tests, plus a parsecache regression test

Collected tests: 1571 -> 1570.

Behaviour changes

• parsecache evicted in the wrong order. The sub-key trim sorted on repr(float(resume_after)) and dropped the lexicographically smallest, so '9.0' outranked '10.0' — recently-used entries were evicted while stale ones survived. Now sorts by recency, as its docstring always claimed.
• RunningSubagent.last_activity tightened from tuple | None to always-tuple. The None branch was unreachable and forced Optional handling at three renderer call sites.

Test defects fixed

• test_token_rate.py: FakeTime._now was mutated and never reset — order-fragile under -n auto.
• test_cohort_visibility.py: stubbed _conclusively_retired to a constant, so the real predicate never ran.
• test_tokens_cost.py: an assertion compared a call to itself and could never fail.

Two pre-existing bugs found, not fixed here

Both are written up in KNOWN_ISSUES.md.

1. Unreachable recovery code, claude/yas/info/subagents.py ~1081-1122. _conclusively_retired requires 125s staleness (max(FINISHED_LINGER, COHORT_GRACE) + TERMINAL_SKEW); visible()'s own _retired() uses 120s on the same branch. Anything conclusively retired is already gone, so the mispredict re-parse block never executes. Pinned as xfail(strict=True). Fix is to align the constants or delete the block.
2. claude/mon/layout.py::_visible_len mishandles wide characters.

Why not more

• themes.py is 1,641 lines of palette data; 0 of 39 Theme fields are identical across all 15 themes.
• The two largest functions, build_wide and subagent_row, are dense column arithmetic, not boilerplate.
• The test tree holds ~1,380 distinct test bodies at a ~5-line floor.
• 412 lines were given back for the performance fix.

Further reduction means deleting tests or working code.

Checklist

• Tests added or updated for new/changed behaviour
• (bug fixes) a regression test now covers the situation so it can't come back
• N/A — no behaviour change, because:

Screenshots / recording

Before	After

• N/A — no visible change

Renderer code changed, but the demo gate shows no drift: a baseline built in a separate worktree and run twice fingerprints the pre-existing elapsed-time noise, and every remaining diff falls in those classes.

Benchmark

Command	Mean [ms]	Min [ms]	Max [ms]	Relative
main	44.9 ± 4.3	42.7	74.2	1.00
PR	45.5 ± 2.4	42.7	58.5	1.01 ± 0.11

• N/A — no performance-relevant change

Converting 16 classes to @dataclass made the branch 1.13-1.18x slower than main, minima non-overlapping. The cost is class-definition time, not per-instance: dataclasses._process_class ran 16 times at import for ~11ms, ~24% of total runtime. The statusline is a cold-start CLI run on every prompt, so import time is runtime. All 16 sites are now hand-written __slots__ classes, at +412 lines. Ruling recorded in CODING_STANDARDS.md and KNOWN_ISSUES.md.

pytest, ruff and the demo gate all stayed green through the regression. Only make bench caught it.

System info

Key	Value
OS	Linux archlinux 7.1.8-arch1-3 #1 SMP PREEMPT_DYNAMIC Tue, 11 Aug 2026 09:16:08 +0000 x86_64 GNU/Linux
Claude Code	2.1.238 (Claude Code)
Terminal	TERM=tmux-256color TERM_PROGRAM=tmux SHELL=/bin/bash COLORTERM=truecolor
Locale	LANG=en_AU.UTF-8 LC_ALL=
Python	Python 3.13.13
uv	uv 0.12.5 (210d1f678 2026-08-14 x86_64-unknown-linux-gnu)