feat(compaction): tiered smart compaction + proactive trigger (M3.2 /…

[wusijian007]

… §1)

Second v3 milestone (§1 Smart Compaction). Upgrades context management from reactive dumb-snip to proactive, tiered, deterministic compaction that cooperates with the M3.1a cache prefix breakpoint. Includes the written §1 design section in docs/v3-kernel-roadmap.md (design-before- code per the roadmap's blast-radius tiering).

M3.2a -- tiered deterministic compaction (context.ts):
compactMessagesTiered() shrinks stale messages IN PLACE instead of
dropping a middle slice (which risks orphaning a tool_use/tool_result
pair -> Anthropic 400). Policy: root task + recent window kept
verbatim; in the stale zone, large tool_result blocks become compact
pointers ("[archived Read(src/x.ts) result: N chars omitted ->
]") and long text is snipped. Token reduction comes
from pointer-izing the whales (tool_results dominate an agent
transcript), so message COUNT is preserved and all pairing stays
valid. Falls back to legacy drop-middle only if shrink isn't enough.
Fully deterministic (no model call) -> eval-safe by construction.

M3.2b -- proactive trigger (query.ts):
At each turn boundary, if the transcript estimate crosses
proactiveCompactionSoftLimitRatio (default 75%) of contextBudgetTokens,
compact down to proactiveCompactionTargetRatio (50%) BEFORE the next
request and yield a compaction LoopEvent. Compacting aggressively
(well below the soft limit) maximizes turns-between-compactions --
each compaction is one cache miss (the amortized cache-reset framing
in the roadmap). The reactive prompt_too_long/max_output path stays
as the safety net. The CLI prints a [compaction] line.

M3.1c -- cache attribution (scoped down, with rationale):
Within a single run the ONLY event that invalidates the rolling
message-prefix cache is a compaction (it rewrites stale messages); a
normal turn merely appends and keeps hitting the cache. fork.ts's
prefixHash is a whole-list hash that changes every appended turn, so
a literal per-turn fork-trace would mislabel every turn as a prefix
miss. Instead we mark the real reset with a query.cache_prefix_reset
profile event at the compaction site. Documented in the roadmap.

Tests: context-budget.test.ts (tiered pointer-ization, root+recent verbatim, archive-once, under-target no-op); query.test.ts (proactive fires past soft limit / stays quiet under it); a 6th eval task "proactive-compaction" that drives a big-file Read past a tiny budget and asserts the compaction event -- the eval gate's deterministic fingerprint updated accordingly (tasks 6, turns 13, in 9600, out 555). CLAUDE.md + roadmap updated; M3.2c (LLM summarizer) stays deferred.

Local: 197 tests, 3/3 green.