0.0.11: the completion budget follows the file, so clangd's ~1 s module-importer answers are not cancelled (UP-25)

[Sunrisepeak]

What and why

Typing nlohmann::js in a module-importing file (qt-demo), completion was slow (~1 s) and sometimes did not come up at all. Root cause measured and split in two:

• clangd (UP-25, issue #24 comment): clangd 23.1 answers every completion on a file that imports modules in 0.95–1.05 s, even when the BMIs are long built and cached — the module context is re-loaded per request (a file without imports in the same project, same CDB and flags: 60 ms; the imports are not in the preamble, so nothing amortizes between requests). This half is upstream.
• mcppls (this PR): the flat completion budget is 1000 ms, so in that regime every engine answer was cancelled right at the line and the fallback answered instead — and after ., -> or :: the fallback is empty by design, so the popup never came.

The change

The budget now follows the file (EnginePace, src/orchestrator/completion.cppm; wired in workspace.cpp):

• The workspace records each core-engine completion answer for a file and its latency (including the late answers C-2 already keeps).
• Two answers in the last ten seconds that all landed past the flat budget extend that file's completion and signature-help budget to their slowest + 500 ms, capped at 2.5 s. qt-demo's importer answers at ~1.05 s: from the third completion on, the engine's real answer lands instead of being cancelled.
• An engine that answers rarely (U16's module rewritten under autosave, U7a's fan-out save) or quickly (ordinary files, 0.3 s) keeps the flat 1 s budget — the fallback never gets slower for anyone.
• The flat budget itself is unchanged from 0.0.10 (routing::answer_budget reverted to main); hover and go-to-definition keep theirs.

An earlier commit here raised the flat budget to 1.5 s; CI's UX gate rejected it with data — U16-module-autosave: 93 completions p95 1.50 s (every one riding to the budget, ux-xlings + ux-mcpp failed, run 37188779178) — which is the measured proof that no flat constant can hold both regimes. The adaptive design passes the same scenarios unchanged.

Registered upstream-defect handling as WA-CLANGD-012 (src/engine/clangd/workarounds.cpp): upstream, evidence, removeWhen, premise. Version 0.0.11 (mcpp run -p devtools -- version --set).

Review pass (1f8083b)

The first adaptive cut had two real holes, found in review and fixed:

• A slow-but-answering engine could push its file's fallback 1.5 s past the flat budget (the regression the gate rejected, narrower): late answers up to 10 s old were all counted, and two 3–8 s answers sent the budget to the 2.5 s cap that those answers would never meet. Answers landing beyond cap − margin (2 s) are not counted — a busy engine is not a slow-and-steady one.
• A think pause reset the learning: the 10 s window cleared on every pause, so the first two completions after pausing missed again — the very symptom being fixed. The window is now 60 s; a longer pause relearns with the answers C-2 already keeps.

Also from review: latency is measured from the ask (where the budget is counted from, too); signature help records its answers instead of only reading completion's; a closed file's pace is dropped; the ring is walked in place; package-lock.json synced to 0.0.11; zh-CN troubleshooting carries the same note as the English page.

Test

• Unit: EnginePace policy (flat until two recent past-the-line answers; fast answer among them → flat; stale window → flat; cap) in tests/test_completion.cpp; R-7 assertions unchanged at 1000 ms.
• Local: mcpp build, mcpp test (all suites, 0 failed), check all, and the completion-keywords conformance fixture against the real payload: 0 failures.
• CI gates the calibrated UX scenarios (ux-xlings, ux-mcpp, ux-heavy-headers) — the first design failed them; this one keeps their numbers.