obsidian-vault/work/projects/cpm-metrics-causes-and-mitigations-2026-10-16.md

title, date, tags, updated

title	date	tags	updated
CPM — metrics, causes and mitigations	2026-10-16	work cpm analysis	2026-10-20

CPM: metrics, causes and mitigations

Updated 20 October 2026. Internal metrics; keep these figures out of public PRs. Production figures retain their original date windows. cpm-e2e-init refers to the historical 16 October test worktree; current PRs and source paths are listed in cpm-status-2026-10-15 and cpm-web-extension-code-guide-2026-10-09.

What the metrics mean

The monitor reports failure when no matching native dashboard refresh arrives within 4 seconds of navigation completion. A second qualifying timeout marks the episode stuck and triggers M4 extension reload. Neither timeout nor recovery proves delivery of initResp to the page: tests can receive it while dashboard delivery fails.

The production cut below uses the old separate daily pixels: installation-days per event/reason, not distinct users or paired episodes. Independently deduplicated failure/recovery totals cannot establish transition or recovery success rates. Keep .count separate from .daily.

The published #7376 branch changes this to one terminal web_extension_cpm_episode, daily+count. Daily counts installations with a completed episode; count counts completed episodes. A new baseline is required. Its current PR is closed without merge, so this is not a shipped-data claim. The four-state stuck and two-phase bg_events contract is documented in cpm-pr0-episode-telemetry-summary-2026-10-16.

Distribution

Latest received production cut: 12–14 October, iOS 7.240+ / macOS 1.210+. Figures below always use iOS / macOS order.

• Initial failures: 3,273,025 / 119,950; stuck: 511,660 / 40,842.
• Recovery without reload: 1,135,477 / 30,738 (34.7% / 25.6% of initial-failure daily totals).
• Recovery after M4: 293,557 / 21,865 (57.4% / 53.5% of stuck daily totals); reload-failed reports: 140,921 / 17,796 (27.5% / 43.6%). These are ratios of event totals, not episode outcomes.

Mutually exclusive snapshot cohorts, in classification priority order. Each pair shows share of initial failures → share of stuck:

• Error 6: 6.1% → 21.5% / 1.0% → 1.9%.
• No background view: 48.1% → 10.9% / 3.3% → 1.8%.
• Tab detached: 0.03% → 0.05% / 6.2% → 8.9%.
• Network process replaced: 16.8% → 39.7% / 13.6% → 26.6%.
• Recent wake: 7.9% → 9.5% / 32.5% → 32.2%.
• Recent context load: 15.7% → 6.0% / 20.0% → 6.4%.
• Other: 5.3% → 12.5% / 23.4% → 22.2%.

These classify observed state, not causes. Network replacement and error 6 are more prominent in stuck; missing-view and recent-load cases are less prominent.

Earlier 5 October post-release cut: initial failures 6,624 / 1,377, stuck 1,440 / 487, first post-reload timeouts 670 / 232. It established the same wake/load distinction; its different window and filters must not be combined with the newer totals. cpm-web-extension-public-attribution-2026-10-05 · cpm-web-extension-data-analysis-summary-2026-10-08

Causes and suggested mitigations

Percentages below are shares of all initial-failure / stuck daily reports in the 12–14 October cut. Snapshot cohorts are proxies for causes; P5/P6 selections use separate denominators.

• Network crash: replacement cohort — iOS init 16.8%, stuck 39.7%; macOS init 13.6%, stuck 26.6%. reproduced across tabs/windows; a surviving background view can retain a dead worker after Network-process replacement. #7459 restarts the background and now reports web_extension_cpm_network_process_recovery_applied (daily+count, no custom parameters). This counts application, not proven recovery or every crash. Production attribution remains incomplete: network_restarted compares against context load, not the latest view. Add network_replaced_since_view; the older analysis supports at most 20–45% of healthy-looking stuck reports, not the whole cohort.
• P1, slow idle wake: recent-wake cohort — iOS init 7.9%, stuck 9.5%; macOS init 32.5%, stuck 32.2%; excludes older wake tails. an imposed startup delay reproduces timeout followed by a late answer; natural production slowness remains unexplained. The published #7379 mitigation prepares background before navigation, with a shared, cancellable 5-second maximum wait. It passed with an 8-second worker delay; longer delays remain possible.
• P2, error 6: error-6 cohort — iOS init 6.1%, stuck 21.5%; macOS init 1.0%, stuck 1.9%. controlled macOS suspension reproduces WebKit's roughly 30-second background-load timeout. The main iOS chains remain unproven. The published #7379 mitigation retries loadBackgroundContent once on the current context, with cooldown and recovery guards. Same-document dashboard and page initResp recovery passed twice. Full-context replacement was rejected after repeat page-response loss. Foreground-resume recovery remains untested.
• P3, loaded context without view: no-view cohort — iOS init 48.1%, stuck 10.9%; macOS init 3.3%, stuck 1.8%. startup dominates on iOS. Base-URL cleanup can delay startup, but the full P3 failure remains unreproduced, including the 20,000-resource cache test. #7493 now implements a stable private extension origin, default ON with a remote kill switch, excluding installations with web-accessible resources. macOS tests verify origin continuity and removal of repeated migration; they still do not reproduce the complete production P3 failure. No new telemetry definitions are included.
• P4, first mid-session load: Fix A signature — iOS init 3.9%, stuck 0.9%. Broader recent-load cohort: iOS init 15.7%, stuck 6.0%; macOS init 20.0%, stuck 6.4%. early init loss is reproduced; #7379 waits for first background readiness and separately counts web_extension_cpm_startup_response_missing with daily+count even while late-document reload is disabled. Late-document reload remains default OFF. The narrower signature is 129k → 4.6k iOS reports and can include reloads.
• P5, background fails to wake: 5 October, after each platform's release tag: iOS init 1.40% (93 / 6,624) → stuck 2.08% (30 / 1,440); macOS init 6.83% (94 / 1,377) → stuck 4.52% (22 / 487). Recounted from the saved cpm-public-attribution-2026-10-06/initialization_daily.tsv and events_daily.tsv: last five bg_events names proxy_on>dealloc>view>proxy_on>dealloc, excluding error 6, Network replacement and scripts=false. This is the earlier 5 October cut, not 12–14 October. 14 October stuck: iOS 1.15% (1,953 / 169,115); macOS 3.57% (499 / 13,992); init for that newer cut has not been recounted. Closing a window with a shared pinned tab reproduces lost tab registration on macOS; fix window/tab lifecycle reporting. That signature appears in 13.6% of macOS P5 (68 / 499), or 0.49% of all macOS stuck. Missing listener registration also reproduces P5 in controlled tests; its natural trigger remains unknown. The earlier 494-report selection used different filters.
• P6, controller mismatch: broader detached-tab cohort — iOS init 0.03%, stuck 0.05%; macOS init 6.2%, stuck 8.9%. P6-only init/stuck shares are unavailable. a main webExtensions flag cycle creates a new manager while existing web views retain the old controller. Recreate/rebind affected views when replacing the manager; context reload alone cannot repair this. The conditional mechanism is reproduced, but its occurrence in production is unconfirmed. Selected macOS tails: 627 + 235 reports.
• Other: iOS init 5.3%, stuck 12.5%; macOS init 23.4%, stuck 22.2%. Tests reproduce slow settings replies, lost messages, and missing dashboard confirmation even when the page receives its answer. Some iOS reports also resemble slow wake-up (P1). Their individual shares are unknown. Mitigations depend on which step fails: wait for background readiness, reduce settings delays, or fix message/confirmation delivery.

PR status and local validation

Verified from GitHub on 20 October 2026; exact heads and CI failure annotations are in cpm-status-2026-10-15#2. PR и quality gates.

• #7376 5d1cb4abd2: closed without merge, approved, CI green. macOS #7484 3fb13b81fe: open, CI green, but release-base conflict and earlier telemetry contract.
• #7379 674bb809f2: open on the telemetry branch; BSK CI failed on a crashed ConfigurationFetcher XCTest. Earlier green results below belong to their tested revisions.
• #7459 f41c7b7047: current Network recovery plus applied pixel; open, CI green.
• #7493 dd96018da9: stable-origin P3 path; open, CI green, full P3 failure still unreproduced.
• #7473 574e9288ae: shared-pinned window-close P5 fix merged 20 October 2026 into release/macos/1.213.0, commit 9bcce0264e; CI green. Unit tests now assert WebKit openTabs; the real-window fixture remains local, with ON 9 → OFF 5 expected failures + 1 control → ON 27 PASS. It fixes one mechanism; release publication, rollout and production effect remain unverified. Details: cpm-status-2026-10-15#P5 — merged 20 октября 2026.
• #7378 harness and #7384 baseline are closed without merge. #7385/native progress and extension #3727 remain open; #7385 still depends on the closed baseline branch. Old #7381 is still open but is not the current Network recovery implementation.

Readiness timing does not establish message delivery. The episode report still lacks page ACK/last confirmed stage, a final result for initially pending loads, and network replacement relative to its current view. Late errors before stuck can now appear in the history continuation; non-stuck reports keep only their initial history. See cpm-status-2026-10-15#6. Телеметрия.

Historical local cpm-e2e-init, 16 October: P1/P2 candidates were uncommitted at that point; the current implementation is in #7379. 22 unique app methods passed across separate runs; two deliberate disabled-fix controls failed. Package validation: 806 tests, 14 skipped, zero failures. Tests use real app/WebKit paths with an instrumented bundle and imposed faults; they do not identify every production cause. P3 remains unreproduced; app P5 establishes one lost navigation. No shipped-effect claim.

Detailed evidence and final validation: cpm-status-2026-10-15 · cpm-web-extension-code-guide-2026-10-09.