Add disk-backed transcript parse cache (4.2x faster subagent-heavy renders)

[tmck-code]

Context

• Statusline renders slow down permanently as a session accumulates subagents.
• each render is a fresh process that re-reads every agent transcript from byte 0, twice (once for the subagent tree, once for tool counts).

Important

This PR implements a cache for a measured speed increase of 318.6 ms > 61.1 ms (5.2x)

Bumps version to 0.8.0

Changes

Transcript parse cache (claude/yas/info/parsecache.py, new)

• Per-session JSON cache at ~/.claude/yas-cache/transcripts.<session_id>.json, version-stamped, atomic os.replace writes, 24 h retention, fail-safe: any corruption/mismatch falls back to a full re-parse.
• Entries keyed by (path, mtime, size) plus parse arguments (resume_after, clear_epoch/skip_sidechain), so argument changes can't return stale results.
• Finished agent transcripts are immutable → warm hit rate approaches 100%; only the main transcript tail and active agents are re-read, incrementally via the existing tail-read machinery.

Cache-backed readers (subagents.py, toolcounts.py)

• parse_transcript, the notification/tool-result tail readers, and count_transcript all seed from and write back to the cache.
• New totals_only parse mode for conclusively-retired agents (cold-cache fallback), with a post-visible() re-parse fail-safe so no rendered row is ever built from a stub.
• ToolCounts.gather still covers every subagent (rendered totals unchanged) — it's just satisfied from the cache.

Lifecycle & config

• Cache loaded once and saved once in app.render(); SessionView keeps its no-disk-writes contract; mon and tests untouched.
• New transcript_cache knob ([cache] in yas.toml, env YAS_TRANSCRIPT_CACHE), default on.

Measured effect (real captured payloads, hyperfine): 48-agent session 318.6 ms → 74.8 ms warm (4.2×); cold render unchanged (311.6 ms); cache file 44 KB. Rendered output is byte-identical cache-on vs cache-off (frozen-clock equivalence tests).

Checklist

• Tests added or updated for new/changed behaviour (+37 tests; suite 1547 passed, ruff + mypy clean)
• (bug fixes) a regression test now covers the situation so it can't come back
• N/A — no behaviour change

Screenshots / recording

• N/A — no visible change (demo/visual gate confirmed zero diff)

Benchmark

Command	Mean [ms]	Min [ms]	Max [ms]	Relative
main	51.7 ± 4.8	42.8	66.4	1.02 ± 0.16
PR	50.6 ± 6.1	44.1	82.8	1.00

(make bench uses a small demo payload, so parity is expected — the gain is on subagent-heavy sessions: 318.6 ms → 74.8 ms warm on a real 48-agent session, measured with hyperfine on captured production payloads.)

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	2.1.233 (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 0.12.3 (507230998 2026-08-07 x86_64-unknown-linux-gnu)

🤖 Generated with Claude Code