fix(sentry): gate non-actionable client noise before it leaves the device (AGE-105)

[dzianisv]

Why

opencode-mobile is now the #1 source of Sentry error volume in the org — ~4,500 events/month against a 3,500/month org quota. Box-bot was silenced by the AGE-55 shim (50.7 -> 0.00 events/h); mobile is what's left.

~1,100 of those events are three non-defects:

462 events / 104 users   Error: connect timeout
498 events /   1 user    Error: API Error: 401 - ...
157 events /  33 users   Error: connect server-unreachable

The connect ... ones are captureDiagnostic() reports of the user's own LAN/VPN/self-hosted box being down. The 401 storm is one device's token-refresh retry loop: 498 copies of one problem, not 498 problems.

What

New pure module src/lib/sentry-noise.ts, wired into beforeSend. Three layers, cheapest first:

Layer	Rule	Rationale
Always-send allowlist	OOM / ANR / native / IllegalStateException / NullPointerException / fatal level / unhandled mechanism	Quota is worthless if it silences real crashes. These bypass everything below.
Transport drop-list	connect timeout/server-unreachable/no-internet/malformed-url, Network request failed, Request timed out after, ECONN*/ETIMEDOUT...	Hard drop, not sampled — the gate runs per-install, so even "1 per device per day" multiplies by the install base back into thousands/month.
Dedup + rate cap	6h per-fingerprint cooldown, <=6 new fingerprints/h, <=10 events/h	Mirrors the openclaw-box-bot shim (AGE-55). Turns a retry loop into one report and caps the blast radius of any future regression.

connect health-failed and connect tls-error are deliberately not dropped — a box that answers but is unhealthy, or a broken cert, is actionable.

We don't lose the signal

Connection failures are already (a) shown to the user as connection UI and (b) counted, PII-free, in PostHog as connection_failed{error_class} via analytics-classify.ts. Sentry was charging us per event for a graph we already have. captureDiagnostic() now short-circuits for those classifications so the event isn't even built.

Drops stay auditable: the count since the last delivered event rides along as a noise.dropped_since_last tag on the next event that does go out.

Verification

• 18 new unit tests in src/lib/sentry-noise.test.ts (pure, deterministic clock — no sleeps).
• Includes a replay of the exact observed hour: 1,126 raw events -> 5 delivered (1 auth report + 4 real OOMs).
• Explicit tests that OOM/ANR/fatal survive after every budget is exhausted.
• npm test: 257 passing. npm run typecheck: clean.

Follow-up (not in this PR)

• Ship in a release build, then re-measure over a clean 7d post-deploy window; target < ~1,500 events/month with genuine crash classes still arriving.
• Investigate why one user produced 498 x 401 — the dedup hides the symptom; there may be a real token-refresh bug underneath.

Tracked in AGE-105.