feat(menubar): say whether a quota window lasts until its reset

[ozymandiashh]

Summary

• The Capacity Dock's window columns and the agent-tab quota hover card now answer "am I going to make it?", not just "when does this reset?": one line under the bar reading Lasts until reset, Runs out in 2d 8h, or Won't last until reset on windows of 6 hours or less, where a linear run-out ETA is not defensible.
• It is the same projection the Plan tab's deficit/reserve caption has used since menubar: quota windows show no pace — add run-out projection (deficit, lasts-to-reset / runs-out-in) #726/menubar: linear pace on Codex quota windows — deficit/reserve, projection, run-out ETA #728 — QuotaPace.evaluate stays the only pace math in the app, and this is a (text, willOverflow) reduction of its result — so the three surfaces cannot disagree. All of menubar: quota windows show no pace — add run-out projection (deficit, lasts-to-reset / runs-out-in) #726's guards come along: silent in the first 3% of a window, on an exhausted window, without a reset time, and on clock skew, with the <= 6h ETA suppression kept rather than crying wolf over one burst.
• QuotaSummary.Window carries no window length, so rather than thread one through ~15 adapter call sites, QuotaPace.inferredWindowSeconds(label:resetsAt:) reads it from the normalized label the adapters already derive from the exact API duration (Weekly, 5-hour, Weekly · Opus, Monthly usage limit). Calendar months are measured back from the reset date in UTC, mirroring how the Codex spend window is derived. A label that implies no length (Copilot's Premium requests, Cursor's Team pool) keeps the row silent instead of guessing.
• The dock reserves the line's height whether or not a column has a verdict, because the glance panel's frame is computed rather than fitted — the same reason the staleness notice has its own reserved height — so a projection crossing in or out of silence cannot resize the bubble under the pointer. The hover card is fitted, so there the line collapses.

Closes #1215

Testing

• I have tested this locally against real data (not just unit tests)
• npm test passes — not applicable, no TypeScript touched; cd mac && swift build succeeds (Build complete, 53.76s)
• npm run build succeeds — not applicable; macOS target only

mac/Tests/CodeBurnMenubarTests/QuotaPaceVerdictTests.swift adds 12 XCTest cases over the pure
(utilization, elapsed fraction, window length) -> String? mapping: all four outcomes
(lasts / runs out with an ETA / won't last on a 5h and on an exactly-6h window / silent), the
early-window silence at 2% and 1% elapsed with 5% speaking, exhausted, missing reset and clock
skew, the countdown wording, the percent 0...1 conversion, and label inference including
Weekly · Opus, GPT-5.3-Codex-Spark · 5-hour, Gemini Models · Five-hour,
Monthly usage limit against a February-in-UTC reset, and the bucket labels that must stay
silent. CapacityDockGlanceTests is updated for the three hard-coded panel heights the
reserved line moves.

Note: my machine has a CLT-only Swift toolchain with no swift-testing module, so the test
target does not compile locally (error: no such module 'Testing' from an existing test file)
and I could not run swift test. To avoid shipping unverified assertions I executed every
expectation in the new file against the real QuotaPace.swift through a standalone swiftc
harness — 43/43 pass — but CI's swift test is the real check here.