fix(usage): apply round-3 review findings to OpenRouter credits

[pandec]

Follow-up to #182. A parallel round-3 review (gpt-5.6-sol read-only + Opus code-reviewer) surfaced six issues in the merged feature; this applies all of them.

Both reviewers independently found the biggest one: the credits query keeps its previous success when a refresh fails, so a disconnected environment kept showing its last balance as current. The settings rows now check the transport error first and say "Unavailable", and the meter dims the number and adds "Couldn't load the latest balance" instead of presenting yesterday's figure.

The rest: key saves and resets from Settings → Extras are chained so they run in click order (the payload-keyed single-flight lanes had made a reset able to race a still-running save and lose); cache TTL checks treat a backwards wall-clock jump as expiry instead of pinning an entry fresh; the popover no longer refreshes the shared empty-query sentinel while the feature is off; and two test gaps are closed — the single-flight gate is now exercised by genuinely concurrent readers, and the hung-body timeout test proves the gate is released afterwards.

vp check and vp run typecheck pass; 12 focused server tests plus the meter and client-runtime suites are green.

Implemented by Claude Fable 5 in Claude Code from review findings by gpt-5.6-sol and Claude Opus.

[pandec]

Follow-up to #182. A parallel round-3 review (gpt-5.6-sol read-only + Opus code-reviewer) surfaced six issues in the merged feature; this applies all of them.

Both reviewers independently found the biggest one: the credits query keeps its previous success when a refresh fails, so a disconnected environment kept showing its last balance as current. The settings rows now check the transport error first and say "Unavailable", and the meter dims the number and adds "Couldn't load the latest balance" instead of presenting yesterday's figure.

The rest: key saves and resets from Settings → Extras are chained so they run in click order (the payload-keyed single-flight lanes had made a reset able to race a still-running save and lose); cache TTL checks treat a backwards wall-clock jump as expiry instead of pinning an entry fresh; the popover no longer refreshes the shared empty-query sentinel while the feature is off; and two test gaps are closed — the single-flight gate is now exercised by genuinely concurrent readers, and the hung-body timeout test proves the gate is released afterwards.

vp check and vp run typecheck pass; 12 focused server tests plus the meter and client-runtime suites are green.

Implemented by Claude Fable 5 in Claude Code from review findings by gpt-5.6-sol and Claude Opus.

[pandec]

Follow-up to #182. A parallel round-3 review (gpt-5.6-sol read-only + Opus code-reviewer) surfaced six issues in the merged feature; this applies all of them.

Both reviewers independently found the biggest one: the credits query keeps its previous success when a refresh fails, so a disconnected environment kept showing its last balance as current. The settings rows now check the transport error first and say "Unavailable", and the meter dims the number and adds "Couldn't load the latest balance" instead of presenting yesterday's figure.

The rest: key saves and resets from Settings → Extras are chained so they run in click order (the payload-keyed single-flight lanes had made a reset able to race a still-running save and lose); cache TTL checks treat a backwards wall-clock jump as expiry instead of pinning an entry fresh; the popover no longer refreshes the shared empty-query sentinel while the feature is off; and two test gaps are closed — the single-flight gate is now exercised by genuinely concurrent readers, and the hung-body timeout test proves the gate is released afterwards.

vp check and vp run typecheck pass; 12 focused server tests plus the meter and client-runtime suites are green.

Implemented by Claude Fable 5 in Claude Code from review findings by gpt-5.6-sol and Claude Opus.

[pandec]

Follow-up to #182. A parallel round-3 review (gpt-5.6-sol read-only + Opus code-reviewer) surfaced six issues in the merged feature; this applies all of them.

Both reviewers independently found the biggest one: the credits query keeps its previous success when a refresh fails, so a disconnected environment kept showing its last balance as current. The settings rows now check the transport error first and say "Unavailable", and the meter dims the number and adds "Couldn't load the latest balance" instead of presenting yesterday's figure.

The rest: key saves and resets from Settings → Extras are chained so they run in click order (the payload-keyed single-flight lanes had made a reset able to race a still-running save and lose); cache TTL checks treat a backwards wall-clock jump as expiry instead of pinning an entry fresh; the popover no longer refreshes the shared empty-query sentinel while the feature is off; and two test gaps are closed — the single-flight gate is now exercised by genuinely concurrent readers, and the hung-body timeout test proves the gate is released afterwards.

vp check and vp run typecheck pass; 12 focused server tests plus the meter and client-runtime suites are green.

Implemented by Claude Fable 5 in Claude Code from review findings by gpt-5.6-sol and Claude Opus.

[pandec]

Follow-up to #182. A parallel round-3 review (gpt-5.6-sol read-only + Opus code-reviewer) surfaced six issues in the merged feature; this applies all of them.

Both reviewers independently found the biggest one: the credits query keeps its previous success when a refresh fails, so a disconnected environment kept showing its last balance as current. The settings rows now check the transport error first and say "Unavailable", and the meter dims the number and adds "Couldn't load the latest balance" instead of presenting yesterday's figure.

The rest: key saves and resets from Settings → Extras are chained so they run in click order (the payload-keyed single-flight lanes had made a reset able to race a still-running save and lose); cache TTL checks treat a backwards wall-clock jump as expiry instead of pinning an entry fresh; the popover no longer refreshes the shared empty-query sentinel while the feature is off; and two test gaps are closed — the single-flight gate is now exercised by genuinely concurrent readers, and the hung-body timeout test proves the gate is released afterwards.

vp check and vp run typecheck pass; 12 focused server tests plus the meter and client-runtime suites are green.

Implemented by Claude Fable 5 in Claude Code from review findings by gpt-5.6-sol and Claude Opus.

[pandec]

Follow-up to #182. A parallel round-3 review (gpt-5.6-sol read-only + Opus code-reviewer) surfaced six issues in the merged feature; this applies all of them.

Both reviewers independently found the biggest one: the credits query keeps its previous success when a refresh fails, so a disconnected environment kept showing its last balance as current. The settings rows now check the transport error first and say "Unavailable", and the meter dims the number and adds "Couldn't load the latest balance" instead of presenting yesterday's figure.

The rest: key saves and resets from Settings → Extras are chained so they run in click order (the payload-keyed single-flight lanes had made a reset able to race a still-running save and lose); cache TTL checks treat a backwards wall-clock jump as expiry instead of pinning an entry fresh; the popover no longer refreshes the shared empty-query sentinel while the feature is off; and two test gaps are closed — the single-flight gate is now exercised by genuinely concurrent readers, and the hung-body timeout test proves the gate is released afterwards.

vp check and vp run typecheck pass; 12 focused server tests plus the meter and client-runtime suites are green.

Implemented by Claude Fable 5 in Claude Code from review findings by gpt-5.6-sol and Claude Opus.

[pandec]

Follow-up to #182. A parallel round-3 review (gpt-5.6-sol read-only + Opus code-reviewer) surfaced six issues in the merged feature; this applies all of them.

Both reviewers independently found the biggest one: the credits query keeps its previous success when a refresh fails, so a disconnected environment kept showing its last balance as current. The settings rows now check the transport error first and say "Unavailable", and the meter dims the number and adds "Couldn't load the latest balance" instead of presenting yesterday's figure.

The rest: key saves and resets from Settings → Extras are chained so they run in click order (the payload-keyed single-flight lanes had made a reset able to race a still-running save and lose); cache TTL checks treat a backwards wall-clock jump as expiry instead of pinning an entry fresh; the popover no longer refreshes the shared empty-query sentinel while the feature is off; and two test gaps are closed — the single-flight gate is now exercised by genuinely concurrent readers, and the hung-body timeout test proves the gate is released afterwards.

vp check and vp run typecheck pass; 12 focused server tests plus the meter and client-runtime suites are green.

Implemented by Claude Fable 5 in Claude Code from review findings by gpt-5.6-sol and Claude Opus.

[pandec]

Follow-up to #182. A parallel round-3 review (gpt-5.6-sol read-only + Opus code-reviewer) surfaced six issues in the merged feature; this applies all of them.

Both reviewers independently found the biggest one: the credits query keeps its previous success when a refresh fails, so a disconnected environment kept showing its last balance as current. The settings rows now check the transport error first and say "Unavailable", and the meter dims the number and adds "Couldn't load the latest balance" instead of presenting yesterday's figure.

The rest: key saves and resets from Settings → Extras are chained so they run in click order (the payload-keyed single-flight lanes had made a reset able to race a still-running save and lose); cache TTL checks treat a backwards wall-clock jump as expiry instead of pinning an entry fresh; the popover no longer refreshes the shared empty-query sentinel while the feature is off; and two test gaps are closed — the single-flight gate is now exercised by genuinely concurrent readers, and the hung-body timeout test proves the gate is released afterwards.

vp check and vp run typecheck pass; 12 focused server tests plus the meter and client-runtime suites are green.

Implemented by Claude Fable 5 in Claude Code from review findings by gpt-5.6-sol and Claude Opus.