feat: surface PHP fatals, disk pressure, and Action Scheduler bloat in wake briefing

[chubes4]

Summary

WakeBriefingTask only scanned three DM-database sources — failed datamachine_jobs, processing datamachine_jobs, and datamachine_logs at level = ERROR. That made it structurally blind to the three infrastructure-layer signals that mattered most in the recent events.extrachill.com incident: a claim_actions PHP fatal storm (fatals never land in datamachine_logs), a 94%-full disk, and a 28M-row / 23GB actionscheduler_actions runaway. The operator caught it via an external fatal alarm, not WAKE.

This wires three new threshold-gated gatherers into gatherSiteSignals() alongside the existing three. Each keeps the ruthless-terseness contract: ONE grouped markdown line, only when it crosses a bar, '' otherwise — exactly like getStuckJobs() / getGroupedErrors().

Signals added

1. PHP fatals from debug.log (getPhpFatals()) — the highest-value add. Reads a bounded tail (5 MiB) of the live debug.log, folds multi-line stack traces into one entry, filters to the rolling window by parsed [DD-Mon-YYYY HH:MM:SS UTC] timestamp, normalizes each fatal/parse-error (warnings/notices excluded) to a stable signature (file basename + numeric-stripped message), groups, and emits e.g. ⚠ 3 PHP fatal(s) in debug.log, top: "foo.php: Uncaught Error: Call to a member function claim_actions() on null" ×2. This mirrors the rotation-safe reading approach behind wp extrachill analytics errors --severity=fatal (extrachill-analytics/inc/core/php-error-log.php) — reimplemented self-contained inside the task rather than reaching across a plugin boundary, so data-machine keeps no dependency on extrachill-analytics.

2. Disk pressure (getDiskPressure()) — disk_free_space() / disk_total_space() on ABSPATH (falls back to WP_CONTENT_DIR, then /). Surfaces when < 15% free OR < 20GB free (whichever triggers). Example: ⚠ Disk 94% full (7.9GB free).

3. Action Scheduler bloat (getActionSchedulerBloat()) — one information_schema.TABLES read of table_rows + (data_length + index_length) for {$wpdb->prefix}actionscheduler_actions and …_logs. Surfaces when > 1,000,000 rows OR > 2GB on either table. Example: ⚠ Action Scheduler bloat: wp_actionscheduler_actions 28.1M rows / 23.9GB.

debug.log path resolution

resolveDebugLogPath() is robust to WP_DEBUG_LOG being a bool or a path string: if it's a non-empty string, use it; else honor a real-file error_log ini target (skipping syslog/stream wrappers); else fall back to WP_CONTENT_DIR/debug.log. If the resolved log is unreadable/absent, getPhpFatals() returns '' — never fatals the task.

Thresholds & filters

Signal	Default	Filters
Disk	< 15% free OR < 20GB free	datamachine_wake_briefing_disk_min_free_pct, datamachine_wake_briefing_disk_min_free_bytes
AS bloat	> 1,000,000 rows OR > 2GB (per table)	datamachine_wake_briefing_as_max_rows, datamachine_wake_briefing_as_max_bytes
Fatal tail scan	5 MiB tail	(constant FATAL_TAIL_BYTES)

All gatherers are fail-soft: any unreadable log, unprobable filesystem, or empty information_schema read returns '' and never throws.

Per-site vs network disk handling

gatherSiteSignals() runs once per blog under the switch_to_blog loop in gatherNetworkSignals(). AS bloat and fatals are genuinely per-site (per-prefix tables, per-site debug.log semantics) and repeat correctly. Disk is host-global, so a naive per-blog emit would repeat the same warning on every site. I added a $disk_emitted instance guard so the disk line is emitted once per run (on the first site that crosses the bar) rather than deduping it into a separate network-only line — this keeps gatherSiteSignals() self-contained and correct in both site and network scope without special-casing the network path.

Tests

New dependency-free smoke test tests/wake-briefing-infra-signals-smoke.php (same convention as retention-action-scheduler-batching-smoke.php) drives all three private gatherers via reflection:

• Fatals: real temp debug.log with in-window grouped fatals, an out-of-window fatal, a multi-line trace, and a warning — asserts grouped count (3), top signature (×2 claim_actions), warning excluded, empty when window excludes all, and fail-soft empty when the log is absent.
• Disk: filter-driven thresholds against the real filesystem — asserts a line when forced over the bar and empty when forced healthy.
• AS bloat: fake $wpdb information_schema read — asserts the 28.1M/23.9GB actions table is named, the within-bounds logs table is not, humanized formatting, empty when both within bounds, the row ceiling trips independently of bytes via filter, and fail-soft empty on a missing table.

Verification

• php -l on both changed files — no syntax errors.
• php -d error_reporting=0 vendor/bin/phpcs inc/Engine/AI/System/Tasks/WakeBriefingTask.php tests/wake-briefing-infra-signals-smoke.php — exit 0.
• php tests/wake-briefing-infra-signals-smoke.php — 13/13 assertions pass.
• php tests/retention-action-scheduler-batching-smoke.php — still passes (no regression).

Closes #2799