fix: stable logged-cost pricing, local-day bucketing, cache savings

[masked8knights]

Follow-up to #17 (Prime Agent usage support). All build on the PR #17 head.

Problems fixed

1. Costs drift downward and don't match the provider dashboard. Prime/Pi were priced from the models.dev catalog (estimated ?? logged), and every 15-minute sync re-priced the same tokens whenever the catalog refreshed (the bundled snapshot and the live models.dev fetch differ 8x for OpenCode Go models). Changed Prime, Pi, and OpenCode to positive-logged-only so the API-logged usage.cost.total from session logs is authoritative and history never changes on refresh.

2. Today's usage shows as yesterday. Days were bucketed in UTC; machines in UTC+10 (or any positive offset) saw local-morning usage land in the previous UTC day. Bucketing now uses the machine's local timezone in the host collector, collectors.ts, historyStartDay(), the dashboard SQL (date('now','localtime',...)), and the app's rangeDays(). Scan cache version bumped so rows re-parse once.

3. Cache savings disappeared for logged-cost agents. cacheSavingsUsd was zeroed when costMode was positive-logged-only. Savings is an estimate from models.dev pricing and is now computed whenever pricing is available, regardless of how the cost itself is derived.

4. "1 agent scan failed" banner for hosted-API users. The OpenCode collector required the local opencode CLI. When the CLI is missing the source is now recorded as skipped (with a clear message) instead of unavailable, so the banner disappears while hosted usage still arrives via Prime Agent sessions.

5. Phantom $0 rows and scan crashes. Zero-token/zero-cost assistant messages are filtered, and a null-usage guard fixes a TypeError that failed the whole host scan; failure messages now include the offending file path.

Verification

• 61/61 vitest tests pass, tsc --noEmit clean, bb plugin build OK.
• Live check on this machine: events now carry pricing_status: logged, day strings match the local calendar, cache savings > 0, source status ready (prime) / skipped (opencode).

[MayankBansal12]

thanks @masked8knights for raising the PR, will take a look and merge it tomorrow

[MayankBansal12]

I resolved the current merge conflicts against main and pushed the resolution in 9d6e7a3.

@masked8knights A few follow-ups before merging:

1. The OpenCode SQL path still buckets and filters by UTC. date(m.time_created / 1000, "unixepoch") and the strftime cutoffs in server.ts do not use the enrolled host local timezone, while parseOpenCode preserves the returned day. This leaves the today-versus-yesterday bug unresolved for OpenCode near timezone boundaries. Please update both the grouping and cutoff and add a boundary regression test.

2. The test suite passes in UTC (65/65), but currently fails under America/Los_Angeles (2 failures) and Pacific/Auckland (1 failure) because several local-day tests assert fixed UTC calendar dates. Please make these fixtures and expectations timezone-safe and add explicit positive/negative offset coverage.

3. Please move the vendored plugin SDK declaration sync into a separate PR. The functional change is relatively small, while the two generated declaration files account for more than 11,000 changed lines and make this bug-fix PR much harder to review.

It would also be useful to make the intended timezone ownership explicit across the plugin server, enrolled hosts, and browser, since the current history boundary and dashboard filters can be evaluated in different timezones.