fix: stabilize system prompt token estimate (#255)

[Beatrice0377]

Summary

ACP already populated state.systemPromptTokens before pruning, but nudge and
acp_status ignored that cache — each re-estimated the system prompt tokens
from a different message view. After compression, the first visible assistant
can belong to a much later turn whose input includes large accumulated history,
inflating the estimate and causing the two numbers to diverge.

Root Cause

• The stable pre-prune cache (state.systemPromptTokens) existed but was never
  consumed.
• nudge re-estimated from the post-prune transform array, where compressed
  messages are removed and the first visible assistant is a late turn.
• acp_status re-estimated from its own fetched message view.
• After compression those views can select different late assistant messages,
  producing inflated and divergent estimates.

Fix

Three minimal production changes:

1. cacheSystemPromptTokens preserves the first stable positive estimate —
   later transforms no longer overwrite it with a degraded post-compression
   estimate (write-if-undefined guard).
2. estimateContextComposition prefers state.systemPromptTokens when it is a
   positive value; falls back to the existing
   estimateSystemPromptTokens(messages) unchanged when undefined.
3. acp_status (collectVisibleMessages) prefers the same cached value; falls
   back to the existing estimate when undefined.

Scope

• No persistence change.
• No model/provider identity invalidation.
• No /acp context change.
• No public API change.
• No version bump.

Tests

• Pure regression tests: degraded post-compression visible history must not
  inflate the system estimate (cached 10K survives first-visible-assistant
  input of ~200K), and undefined cache falls back to the original estimate.
• acp_status overview regression asserting the cached value appears in the
  breakdown.
• §5.7 multi-turn growth-cycle test (preserveRecentMessages: 20): baseline →
  growth → nudge → compress → new baseline → growth → nudge again, asserting the
  stable system estimate survives compression.
• Docker E2E: autonomous nudge → compress → growth → re-nudge with a verifier
  that ≥2 nudge observations report the same system prompt estimate.
• Bug-targeting tests verified to FAIL with the fix reverted and PASS with it
  applied.
• Verified: npm run typecheck, npm run build, npm run test (1010 pass),
  Docker E2E (12/12 scenarios), check-pr.sh.
• Live OpenCode smoke: plugin loads, ACP state initializes, acp_status
  executes without ACP errors. This verifies the runtime path only — the precise
  estimateSystemPromptTokens inconsistent between nudge and acp_status after compression #255 regression is covered by the automated tests; the live environment was
  not used to reconstruct a 200K post-compression session.

Known limitations

• The cache is in-session only; a restart after host compaction may re-estimate
  from degraded history.
• Mid-session instruction/AGENTS.md changes do not invalidate the cache.

Both are intentionally out of scope for this fix.

Fixes #255

[Beatrice0377]

Dual-agent review

Independent review of the current PR diff is complete.

• Reviewer A: APPROVE — no blocking or non-blocking code findings.
• Reviewer B: APPROVE — no blocking or non-blocking code findings.
• Source changes, regression tests, §5.7 multi-turn coverage, and Docker E2E changes were reviewed independently.
• No additional code changes are requested from the reviews.

CI note

All code-validation checks are passing:

• CI / build ✅
• CI / test (Node 22) ✅
• CI / test (Node 24) ✅
• CI / e2e ✅
• PR Checks / pr-validation ✅

The only failing required check is PR Build Artifact / build-artifact.

Its checkout, npm ci, and npm run build steps succeed; it fails specifically at npm publish with ENEEDAUTH because NODE_AUTH_TOKEN is empty for this fork PR. I have not modified the workflow, package version, or npm authentication to work around this.

Leaving the artifact check for maintainer-side handling.