[BUG] Regression v2.1.74: ~3% quota consumed per 20k token session — Max20 plan

Plan: Claude Max20 (£200/month)
Version: v2.1.74
Models tested: Sonnet 4.6, Opus 4.6 (medium thinking)
OS: Windows 11

Description:

Quota is depleting at a rate inconsistent with actual token usage across multiple independent sessions. This appears to be a regression — #22435 was marked fixed in v2.1.62 but the behaviour persists on v2.1.74.

Data collected (three sessions, same 5-hour window):

| Session | Model | CC-reported tokens | Usage delta |
|---|---|---|---|
| 1 | Opus 4.6 | 42k | +6% (69→75%) |
| 2 | Opus 4.6 | 22k | +3% (75→78%) |
| 3 | Sonnet 4.6 | 20k | +3% (78→81%) |

Implied window limit: ~667k tokens. At this rate, the 5-hour window is exhausted in approximately 6 average sessions.

Expected: Usage proportional to actual token consumption, consistent with a Max20 subscription.

References: #29000, #22435