[R1] Only return QuotaExhaustedError when quota filtering caused empty
candidates (ProviderQuotaSelector.FilteredCount > 0), preventing wrong
HTTP 503 for non-quota failures like model-not-found.

[D1] Extract writeQuotaExhaustedResponse helper, replacing 5 identical
copy-pasted quota error blocks across chat.go, doubao.go, openai.go.

[D2] Remove unreachable quota-exhausted checks after ReadHTTPRequest
in doubao.go and openai.go (ReadHTTPRequest only returns io.ReadAll
errors, never QuotaExhaustedError).

[R5] Rewrite quota test to exercise actual production code
(writeQuotaExhaustedResponse) instead of manually reimplementing
handler logic that would pass even if handler broke.

[S1] Define QuotaEnforcementMode typed constants in backend
(ExhaustedOnly, DePrioritize), replacing 40+ raw string literals
across system.go, orchestrator, resolvers, and tests.

[S1] Type QuotaChannelStatus.Status as providerquotastatus.Status
instead of bare string, reusing existing typed enum.

[S2] Extract hardcoded 0.8 warning usage ratio into named constant
(warningUsageRatio) in lb_strategy_quota.go.

[S1] Define QuotaEnforcementMode union type in frontend (TypeScript),
applied to QuotaEnforcementSettings and UpdateQuotaEnforcementSettingsInput.

[S3] Remove dead 'exhausted_only' fallback and 'value &&' guard in
quota-settings.tsx Select component.

Negate de-prioritize score so warning channels rank below exhausted
ones. Add validation rejecting invalid quota enforcement modes.

Fix playground error handling: return 503 for quota-exhausted errors
instead of falling through to generic HTTP error handling. Update
quota enforcement settings resolver to use the non-default variant
so errors surface instead of silently returning zero values.

Clean up stale .playwright-cli/ test artifacts and add to .gitignore.

Populate per-limit QuotaLimitStatus in all 7 checkers based on request modality
(req.Image != nil for image vs token types). Add NextCheckAt =
checkInterval/4 for warning channels. Propagate quota limit type via
context keys. Score and filter providers per-limit using per-limit
effective status evaluation.

QuotaEnforcementMode was missing MarshalGQL causing gqlgen runtime type
assertion error when serializing the mode field. Converted mode from
GraphQL String to proper enum type with MarshalGQL/UnmarshalGQL methods
and updated frontend to use SCREAMING_SNAKE_CASE values.

…t EffectiveStatus init

- ProviderQuotaSelector: skip filtering in DePrioritize mode so
  QuotaAwareStrategy can penalize exhausted channels instead of
  removing them, preserving the soft enforcement semantic
- EffectiveStatus: initialize worstStatus to StatusUnknown and
  worstReady to false so that Unknown-ranked limits can be
  overridden by any known status instead of returning Available
  incorrectly
- Start: run loadQuotaCache async to avoid blocking startup on
  large quota datasets
- Update tests to expect all candidates in DePrioritize mode

Reduce the blind spot between quota checks so that channels in warning
state are detected sooner. At 20m, a Codex channel at 90% usage on the
3h primary window could fully exhaust before the next check. At 5m,
the worst-case blind spot is under 30% of remaining quota even for the
shortest provider windows.

The adaptive warning interval (base/4) now checks every ~75s instead of
5m, catching rapid exhaustion quickly once warning is first detected.

Also enable refetchIntervalInBackground on the frontend quota query so
the UI stays current when the browser tab is backgrounded.

Add channel-level EffectiveStatus floor when base is exhausted
- Fallback to "unknown" for non-standard status values for consistency
- Lower synthetic warning boundary to prevent false 429 selector picks
- Refactor quotaSelector for correct composition order
- Add UnmarshalGQL for EffectiveStatus type consistency
- Add round-trip test for quota selection logic
- Add config documentation for quota warning interval ratio

Refactor API-specific error formatting for quota exhausted
errors (extract wrapQuotaExhaustedAsResponseError shared helper).

Fix provider_quota warning interval math (multiply vs divide ratio).
Add DE_PRIORITIZE exhaustion detection in candidate selection.
Propagate DB errors in UpdateQuotaEnforcementSettings resolver.
Add usage ratio to wafer quota limits.
Extend synthetic checker test coverage for limits array.
Fix comment: de_prioritize description.

Replace magic numbers (0.8, quota_exhausted) with named constants.
Extract NewTokenLimitStatus() and IsReadyStatus() helpers to deduplicate
quota data construction across 6 provider checkers. Tighten
nextCheckIntervalForStatus() to use providerquotastatus.Status instead
of string. Simplify quota selector wiring in selectCandidates().

Restore doc comments and inline field comments on QuotaChecker interface, QuotaData struct fields, and NanoGPT checker types/functions that existed on unstable but were dropped during the provider_quota package restructure.

Restore doc comments on buildNanoGPTQuotaURL and findEarliestResetAt, plus inline comments on warning state check, nextResetAt calculation, and grace period fallback that were lost during the provider_quota restructure.

…ments

…dges

Rename EXHAUSTED_ONLY enforcement mode to Block Exhausted in UI and locale
files for clarity. Add enforcement effect badges (Blocked/Deprioritized)
to quota dialog rows when a channel's quota is exhausted.

Changes:
- QuotaRow now receives enforcementMode prop and renders Blocked or
  Deprioritized badges based on quota status and enforcement setting
- Update EXHAUSTED_ONLY label in en/zh-CN locale files
- Add quota.status.blocked and quota.status.deprioritized keys
- Minor formatting and import cleanup in quota-badges.tsx