Capture lifecycle and budgets
The capture budget is part of the feature, not an optional optimization.
Current browser lifecycle
Not loaded → authorized explicit on → collecting
→ dump / analyze (user actions)
→ off / TTL / pagehide / failure → inert until reload
The entry module checks the grant before dynamically importing collection code. The collector repeats the origin/environment check. Off during a pending import cancels startup. Once started, a session cannot be restarted in the same document.
On stop it cancels rAF/timers, disconnects observers, closes BroadcastChannel, removes the overlay, clears typed/object buffers and memory accessors, stops the optional sampler and aborts direct provider analysis. Browser memory measurement has no cancellation API; an outstanding result is ignored and no new call starts.
| Current browser cap | Value |
|---|---|
| TTL | Default/maximum 300 seconds |
| Retained frame records | 32,768 |
| Rolling median | 120 frames; at least 10 for spike warm-up |
| Metric buckets / observed memory grows | 300 each |
| Observer intervals | 512 |
| Optional function samples | 10,000; interval 10–1000ms |
| Optional heap/document measurement | At most once per 30 seconds; one pending call |
| Dump | 50MiB total, 1MiB per line |
| Claude summary / response | 24,000 UTF-8 bytes / 256KiB |
| Claude timeout | 15 seconds |
Overwrites are counted. A dump represents retained buffers; a long high-refresh capture can lose its oldest frames before TTL. That must not be presented as an untruncated recording. Explicit export/analysis allocates temporary output; the no-allocation constraint applies to the hot rAF callback.
Current native proof
The separate local dashboard now has an explicit viewer session:
default/maximum TTL 300s, no process reads until Start, at most 300 live buckets,
four WebSocket clients and eight in-flight HTTP handlers. End session, TTL,
signals or target exit closes it and flushes native data subject to a 200MiB
storage cap. Its /proc collection is independent of the BPF proof below.
One selected uprobe and bounded per-CPU hit counters, read at most 1Hz. Default TTL is 30 seconds, maximum 300. RAII closes the link before program/map objects; kernel FD cleanup handles forced termination. No listener, ring buffer or trace writer is created by this proof.
Planned native budgets
Lite sampling: 49Hz. Full: 99Hz. High-frequency sched/syscall data is aggregated in kernel, not streamed per event. Ring buffer cap: 4MiB with counted drops; dump cap: 50MiB. Planned governor reduces collection when the measured budget is exceeded.
Targets, not measured claims: frame/CPU/memory-only under 0.5%, Lite under 1%, Full under 3%. Validate these with the actual sample game and target hardware before stating them as achieved performance.