fix(auth): keep failure bookkeeping when cooling is disabled

[warelik]

Summary

This PR fixes a bug where disable-cooling: true erased all evidence of 429/quota and rate limit failures instead of merely skipping the blackout window. It preserves failure bookkeeping (BackoffLevel increment and UpdatedAt timestamp) under disableCooling while maintaining the flag's contract that credentials stay immediately usable without blackout or suspension. It also adds per-attempt failure logging to MarkResult.

Production symptom

Under disable-cooling: true, when rate limits were encountered:

• The proxy walked all 12 candidates and looped every ~2 minutes against rate-limited credentials.
• A grep for cooldown|quota|backoff|suspend in the server log returned zero lines because failures were completely unrecorded and unlogged.

Root cause and lines that erased state

1. In sdk/cliproxy/auth/conductor_cooldown.go, under MarkResult case 429, quotaCooldownAfterFailure was skipped when disableCooling was true, preventing BackoffLevel from incrementing.
2. In updateAggregatedAvailability, when state.Quota.Exceeded was false (reset by disableCooling), auth.Quota.BackoffLevel was unconditionally reset to 0.
3. In applyAuthFailureState, 429 handling did not advance auth.Quota.BackoffLevel when disableCooling was true.
4. MarkResult did not emit any log line when an execution attempt failed, making repeated failure loops invisible.

Changes

1. Failure bookkeeping under disableCooling:
   • In MarkResult (case 429) and applyAuthFailureState, advance BackoffLevel via nextQuotaCooldown(prevLevel, false) when disableCooling is true.
   • In updateAggregatedAvailability, aggregate maxBackoffLevel across model states and assign auth.Quota.BackoffLevel = maxBackoffLevel even when Quota.Exceeded is false.
   • Record the failure timestamp on state.UpdatedAt and auth.UpdatedAt.
2. Contract preservation:
   • The contract of disable-cooling: true is strictly preserved: Unavailable remains false, Quota.Exceeded remains false, NextRetryAfter remains zero, NextRecoverAt remains zero, no model suspension is triggered, and the credential remains immediately usable without blackout.
3. Per-attempt failure logging:
   • Added structured single-line warning log in MarkResult on every failed attempt using the house logging format:
     auth-cooldown: attempt failed | auth=<id> status=<status> class=<class> cooldown=<duration|skipped> backoff=<level> disable_cooling=<bool>

Sample log line

time="2026-08-21T07:13:27+03:00" level=warning msg="auth-cooldown: attempt failed | auth=auth-1 status=429 class=quota cooldown=skipped backoff=1 disable_cooling=true"

Tests

• TestDisableCoolingRecordsBackoffAndTimestampWhileStayingUsable
• TestDisableCoolingDisabledKeepsStandardCooldownBehavior
• TestMarkResultPerAttemptFailureLogging

Verbatim bite-check failure

Reverting the BackoffLevel increment under disableCooling produced the expected failure:

=== RUN   TestDisableCoolingRecordsBackoffAndTimestampWhileStayingUsable
time="2026-08-21T07:14:01+03:00" level=warning msg="auth-cooldown: attempt failed | auth=auth-disable-cooling-records status=429 class=quota cooldown=skipped backoff=0 disable_cooling=true"
    cooldown_backoff_test.go:354: expected BackoffLevel=1 after first failure, got 0
--- FAIL: TestDisableCoolingRecordsBackoffAndTimestampWhileStayingUsable (0.00s)
FAIL
FAIL	github.com/router-for-me/CLIProxyAPI/v7/sdk/cliproxy/auth	0.457s
FAIL

Verbatim green test output

=== RUN   TestDisableCoolingRecordsBackoffAndTimestampWhileStayingUsable
time="2026-08-21T07:13:27+03:00" level=warning msg="auth-cooldown: attempt failed | auth=auth-disable-cooling-records status=429 class=quota cooldown=skipped backoff=1 disable_cooling=true"
time="2026-08-21T07:13:27+03:00" level=warning msg="auth-cooldown: attempt failed | auth=auth-disable-cooling-records status=429 class=quota cooldown=skipped backoff=2 disable_cooling=true"
--- PASS: TestDisableCoolingRecordsBackoffAndTimestampWhileStayingUsable (0.00s)
=== RUN   TestDisableCoolingDisabledKeepsStandardCooldownBehavior
time="2026-08-21T07:13:27+03:00" level=warning msg="auth-cooldown: attempt failed | auth=auth-normal-cooling status=429 class=quota cooldown=1s backoff=1 disable_cooling=false"
--- PASS: TestDisableCoolingDisabledKeepsStandardCooldownBehavior (0.00s)
=== RUN   TestMarkResultPerAttemptFailureLogging
time="2026-08-21T07:13:27+03:00" level=warning msg="auth-cooldown: attempt failed | auth=auth-log-test status=429 class=quota cooldown=skipped backoff=1 disable_cooling=true"
--- PASS: TestMarkResultPerAttemptFailureLogging (0.00s)
PASS
ok  	github.com/router-for-me/CLIProxyAPI/v7/sdk/cliproxy/auth	10.310s

Tooling note

The jbcontext CLI is installed on this machine but its stored session cannot be decrypted (the OS keychain is not accessible), so jbcontext search could not run. The equivalent semantic search, review and blast-radius passes were performed with local code-intelligence tooling instead.