[Bug][0.1.7-rc.2] Failed turn (HTTP 400/429) leaves all-zero usage samples; contextPressure projection permanently zeroed — ContextMeter shows 0%

[Phantomcyber-ai]

Summary

When a turn fails with a provider error (HTTP 400/429), a placeholder all-zero usage chunk
({ inputTokens: 0, outputTokens: 0, totalTokens: 0 }) is written into the assistant/attempt
stream. The contextPressure projection in dsh-token-meter folds this zero sample as if it were
a valid measurement and overwrites the last good pressureTokens value. Since the
projection-cache checkpoint sits at the session tail, the zero is never recomputed on later loads —
the UI ContextMeter permanently shows 0% for that session, even though the session data
itself is intact.

Reproduction

1. Session with a 1M-token context window; last valid usage sample: input 399 + cacheRead 274560
   → pressureTokens = 274959 (ContextMeter ≈ 27%).
2. Submit a turn; the provider fails. Observed with both 400: insufficient credits (single
   attempt) and 429: token plan entitlement exhausted (retry loop, 5 attempts, one all-zero
   usage chunk per attempt, turn/end error).
3. Reopen the session: the persisted projection checkpoint now holds
   contextPressure: { pressureTokens: 0, sampledSurfaceTokens: <n> }; projected tokens evaluate
   to max(0, …) = 0 → ContextMeter shows 0% permanently (checkpoint seq equals the session
   tail, so replay never recomputes it).

Current behavior

• ContextMeter shows 0% forever after a failed turn; tokenUsage.last is polluted with all-zero
  buckets the same way (per-session totals unaffected because the turn ended with the error).
• Any projection-cache re-fold re-poisons the checkpoint, because the zero samples live in the
  persisted event log (v4 JSONL) — e.g. a cache rebuild / version migration restores the 0%.
• Observed on 2 of 11 v4-format sessions on this machine — exactly the two whose turns ended in
  provider errors.
• Cross-check: the session JSONL itself is intact (final assistant/message usage shows
  totalTokens = 275685), and offline replay of the same raw events through the real kernel
  packages (Session.create + token-meter measure) returns totalTokens = 277806, baseline
  usage — the meter computes correctly from events; only the persisted projection checkpoint
  got poisoned.
• No data loss in our case: the checkpoint was repaired by restoring pressureTokens to the last
  valid value (input + cacheRead of the last valid sample), after which the meter displays
  correctly again.

Expected behavior

Usage samples belonging to failed attempts (HTTP 4xx/5xx, turn ended with error) must not be
folded into contextPressure. Defensively, an all-zero sample (totalTokens === 0 && inputTokens === 0 && outputTokens === 0) should be ignored whenever a previous valid sample
exists.

Environment

• DeepSeek Harness Desktop 0.1.7-rc.2 (exe FileVersion 0.1.7-rc.2), Windows 11 (build 26200)
• Kernel packages @deepseek-ai/* all 0.1.7-rc.2; affected package @deepseek-ai/dsh-token-meter
  (repo path packages/llm/token-meter)
• Session format: v4