fix(ai): Command Code quota monthly window vanishes when remaining credits exceed the hard-coded 10 included — an over-limit payload reads as "could not be parsed"

Verified 2026-11-05 on origin/main fba8bcca0b (laptop primary, clean), probe at ~/git/fleet/probe-2026-11-05-creator-opencode-0815/probe.ts on the laptop (bun, run against the real parser).

The bug

packages/ai/src/quota/providers/command-code.ts:62-72 — parseCommandCodeCredits only publishes the monthly window when the remaining balance falls inside the assumed included range:

export const MONTHLY_INCLUDED_CREDITS = 10
const monthlyRemaining = toNumber(credits?.monthlyCredits)
if (monthlyRemaining !== null && monthlyRemaining >= 0 && monthlyRemaining <= MONTHLY_INCLUDED_CREDITS) {
  const monthlyUsed = MONTHLY_INCLUDED_CREDITS - monthlyRemaining
  windows.monthly = toUsageWindow({ usedPercent: ..., valueLabel: `${monthlyUsed} / 10`, ... })
}

Any remaining balance above the hard-coded 10 (a plan change, a promo/rollover grant, a higher included tier, an upstream field whose meaning is total rather than remaining) drops the window entirely — no window, no error, no clamp.

Reproduced (probe, payload-shaped, against the real parser)

• parseCommandCodeCredits({ credits: { monthlyCredits: 5 } }) → monthly window, usedPercent: 50, label "5 / 10" ✓
• parseCommandCodeCredits({ credits: { monthlyCredits: 10 } }) → monthly window, usedPercent: 0 ✓
• parseCommandCodeCredits({ credits: { monthlyCredits: 25 } }) → {} — no windows at all
• parseCommandCodeCredits({ credits: { monthlyCredits: 25 }, windowLimits: { fiveHour: { used: 1, cap: 4, resetAt: 1893456000 } } }) → only the 5h window; monthly silently absent
• Downstream: when the monthly credit is the only data in the payload, fetchCommandCodeUsage (command-code.ts) throws "Command Code usage data could not be parsed" on a perfectly parseable payload, and the provider reports a hard failure instead of a healthy 0%-used month.

Why it matters

This is the same failure shape the repo just fixed for Turso in #1743 ("publish over-limit windows above 100 percent", merged PR #1743): a value outside the assumed range must be published truthfully (here: 0% used, with the real remaining figure visible), never dropped. A dropped monthly window reads downstream as "no quota data" — the opencode quota display loses the provider's headline window and quota-aware consumers reason from an absence. The hard-coded MONTHLY_INCLUDED_CREDITS = 10 makes the trigger more likely over time, not less: the first plan that includes more than 10 monthly credits breaks the read.

Negative remaining (monthlyCredits: -3) also drops the window in the probe; that is nonsense data and dropping it is defensible — the acceptance below only requires the over-included case to publish.

Acceptance criteria

• parseCommandCodeCredits with monthlyCredits above MONTHLY_INCLUDED_CREDITS publishes a monthly window with usedPercent: 0 and a value label that shows the real remaining figure (not a fabricated "0 / 10" that hides it).
• In-range behavior is unchanged: 5 → 50% used, label "5 / 10"; 10 → 0% used.
• A payload whose only data is an over-included monthly credit no longer makes fetchCommandCodeUsage throw "could not be parsed".
• Tests in packages/ai/test/quota.test.ts (or a wired sibling) pin the over-included case alongside the in-range cases.
• Existing quota tests still pass under the ai package test command; ai typecheck clean.