fix paper link

Opus subagent analysis of v4.1 baseline (333 blocks) vs v4.2 stubs (689
blocks) at the same 258K-token budget recommended four mitigations to
address moderate-risk findings:

1. Recency cue [t-NNm] on turn headers
2. Semantic stub wrapping <lcm-stub> XML tags
3. Empty-assistant collapsing
4. Resolution markers at completion boundaries

Applied first-principles-architectural-decision skill (research,
run-the-system, where-it-lives diagrams, adversarial debate) before
building any of them. Verdict: REJECT ALL FOUR. Each fails on a
specific load-bearing constraint:

- #1 fails on prefix-cache stability (clock-based tag changes the
  rendered string on every assemble, invalidating the cache that v4.2's
  whole value proposition relies on). User timestamps already exist
  inline.

- #2 fails on "novelty has cost, format already works" — the existing
  [LCM Tool Output: file_xxx | …] bracket form is correctly parsed by
  Opus in live tests (drilldown via lcm_describe works on Option F
  format). Replacing a working v4.1-trained format with a novel XML
  form is unjustified churn.

- #3 fails on Anthropic/OpenAI wire contract. The "empty assistants"
  contain tool_use blocks (required to live in assistant turns; paired
  with tool_results by toolCallId). Dropping them would break
  pairing — providers reject orphan tool_results.

- Martian-Engineering#4 fails on detection signal. No reliable way to mark "work
  completed" — user phrases like "go ahead" / "yes" / "keep digging"
  oscillate. False positives are strictly worse than no marker
  (license premature stubbing).

Adversarial debate at ≥95% confidence target on each. AGAINST won on
all four. Decision record committed for future operators who hit
similar moderate-risk findings and reach for similar mitigations.

Final v4.2 shipping shape: Options C + D + F at commit e309bed.
Architecturally additive, reversible, default-off. Empirically:
333→689 items at same budget; Opus drills down correctly; no
confabulation observed.

Wire #1 of 3 for the agent context-management architecture (Wave-14).

# What this lands

Subscribes to the openclaw `llm_output` hook to maintain a per-session
cache of (currentTokenCount, tokenBudget). Anchors via llm_output;
extends via per-tool additive self-updates between LLM calls so
parallel-tool-call sequences see accurate cumulative state, not stale
ground-truth from the previous LLM call.

Replaces lcm_compact's previously-vapor `getRuntimeContext` callback
with the real cache. Floor-check (currentRatio < reserveFraction)
finally works against live data instead of always-undefined.

# Files

NEW:
- `src/plugin/token-state.ts` (~165 LOC) — recordLlmOutput +
  accumulateToolResultTokens + getRuntimeContext + inferTokenBudget
- `test/v41-token-state.test.ts` (~110 LOC) — 16 tests covering
  anchoring, accumulation, drift reset, session isolation, budget
  inference

EDITED:
- `src/plugin/index.ts` — `api.on("llm_output", ...)` handler that
  records tokens; lcm_compact registration passes
  `getRuntimeContext: () => getTokenStateRuntimeContext(ctx.sessionKey)`
- `src/tools/lcm-compact-tool.ts` — docstring on getRuntimeContext
  documents the wiring + tolerance for undefined fields

# How it works

Round 1 (anchor):
  LLM call N → llm_output fires → recordLlmOutput stores
  { currentTokenCount, tokenBudget, lastUpdateSource: "llm_output" }
  keyed by sessionKey

Round 2 (parallel-tool-call protection):
  Tools fire sequentially between LLM calls. Each tool's execute()
  ends with accumulateToolResultTokens(sessionKey, resultText) which
  adds Math.ceil(resultText.length / 4) to currentTokenCount. This
  way a 5-tool batch from one LLM response sees accurate cumulative
  state at each tool, not the same stale value.

Round 3 (drift reset):
  Next LLM call → llm_output snaps cache back to ground truth.
  Any per-tool estimation drift bounded by one iteration's batch.

# Why this layer (vs. waiting for openclaw SDK addition)

`OpenClawPluginToolContext` does not expose token state today. Wave-14
research confirmed (lossless-claw#472, openclaw#68930 closed
NOT_PLANNED). The proper fix is an openclaw PR adding
`getTokenState?: () => TokenSnapshot` to the factory context. That PR
will be filed separately.

This module is the LCM-side bridge that makes the architecture work
TODAY without openclaw changes. Once openclaw lands the official
accessor, this hook handler becomes legacy / fallback for older
versions and the per-tool accumulator stays as a within-iteration
lag-protection layer.

# Verification

- 1573/1573 tests passing (1557 baseline + 16 new)
- 7/7 release-readiness preflight checks pass
- 330 TS errors (under 700 baseline; PR introduced none)

Opus subagent analysis of v4.1 baseline (333 blocks) vs v4.2 stubs (689
blocks) at the same 258K-token budget recommended four mitigations to
address moderate-risk findings:

1. Recency cue [t-NNm] on turn headers
2. Semantic stub wrapping <lcm-stub> XML tags
3. Empty-assistant collapsing
4. Resolution markers at completion boundaries

Applied first-principles-architectural-decision skill (research,
run-the-system, where-it-lives diagrams, adversarial debate) before
building any of them. Verdict: REJECT ALL FOUR. Each fails on a
specific load-bearing constraint:

- #1 fails on prefix-cache stability (clock-based tag changes the
  rendered string on every assemble, invalidating the cache that v4.2's
  whole value proposition relies on). User timestamps already exist
  inline.

- #2 fails on "novelty has cost, format already works" — the existing
  [LCM Tool Output: file_xxx | …] bracket form is correctly parsed by
  Opus in live tests (drilldown via lcm_describe works on Option F
  format). Replacing a working v4.1-trained format with a novel XML
  form is unjustified churn.

- #3 fails on Anthropic/OpenAI wire contract. The "empty assistants"
  contain tool_use blocks (required to live in assistant turns; paired
  with tool_results by toolCallId). Dropping them would break
  pairing — providers reject orphan tool_results.

- Martian-Engineering#4 fails on detection signal. No reliable way to mark "work
  completed" — user phrases like "go ahead" / "yes" / "keep digging"
  oscillate. False positives are strictly worse than no marker
  (license premature stubbing).

Adversarial debate at ≥95% confidence target on each. AGAINST won on
all four. Decision record committed for future operators who hit
similar moderate-risk findings and reach for similar mitigations.

Final v4.2 shipping shape: Options C + D + F at commit e309bed.
Architecturally additive, reversible, default-off. Empirically:
333→689 items at same budget; Opus drills down correctly; no
confabulation observed.