Release 0.3.1: fix Grok usage after period resets

[thrr87]

Grok Build can return a valid new usage period after reset without creditUsagePercent. Codex Limits previously rejected that response, retained the expired snapshot, and reported that the account had no allowance.

Accept the period-only reply as a partial snapshot: retain reset, plan, balance and receive time while leaving the percentage unavailable. Do not insert a zero/100% observation or forecast. Preserve historical points, label their window as “Last recorded”, and resume numeric observations when Grok supplies them. Explicit null/malformed percentages and unusable legacy limits remain rejected. Prepare version 0.3.1 (build 11).

Validation:

• Regression test failed with missingAllowance before the fix and passes after it.
• Full Release suite: 646 tests, zero failures; reset → partial reply → relaunch → explicit-zero recovery is covered.
• Live installed Grok Build 1.0.30: partial current-week reply accepted; cache round trip passed; no model request or login flow.
• Isolated signed native QA: All/detail show unavailable percentage, current reset and retained history; visible copy checked.
• Release-validator self-test and graceful/forced Grok process cleanup passed.
• Spec and strict maintainability reviews found no actionable issue in the Grok change.

The first full local run exposed an existing Codex timer race in testDisplayBoundaryExpiresCodexWithoutAnotherSourceRead (its reset is only 150 ms after fixture creation). The isolated baseline test and the subsequent full 646-test run passed. Codex collection/timer logic is outside this change.

Release authorization (2026-09-16): the product owner explicitly renewed the all-enabled idle comparison and eight-hour mixed lifecycle soak exceptions specifically for 0.3.1; both remain unperformed, not passed. The owner also authorized an administrative merge of PR #74 without a second/code-owner approval after tests and CodeQL pass, followed by publication. ADR-0015 and the PRD record these exact-version exceptions.

The initial implementation head passed GitHub CI (646 tests and process-cleanup checks): https://github.com/thrr87/codex-limits/actions/runs/35109868766. Final head 8e888c2d68ceb359d9b5b0b2fd2c6db5011e590d passed CI (646 tests) and CodeQL for Swift and Actions. Signed universal dry-run and draft artifacts must be verified before publication. Claude Code remains experimental, and its real Pro/Max allowance path remains unverified.