Cross-platform capture smoke evidence

This directory holds the schema, a canonical sample, and the manual evidence policy for the production-path capture fidelity smoke. The smoke proves that the current Krometrail capture pipeline (ProductionBrowserConnector + cdpkit 0.4.0 + supervised targets + bounded ingestion) captures frames, reports loss honestly, and cleans up after itself on supported platform/Chrome combinations.

Files

• schema.json — Draft 2020-12 JSON Schema for CrossPlatformSmokeEvidence.
• sample.json — Hand-authored, schema-valid canonical example pinned to the serializer's canonical byte layout. The deterministic round-trip test asserts that serde_json::to_vec_pretty(&CrossPlatformSmokeEvidence::sample()) equals this file byte-for-byte.
• linux-chrome.json — Decisive Linux Chrome evidence from revision 595f079ce9772e3974f65bdc62ad70b01c17dbce.
• macos-chrome-default-dpi.json — Decisive macOS default-DPI evidence from revision 4dfc78b2a6e91c3404fc58c3c8b98c5b6d662fdc and hosted run 29288634536.
• README.md — This file: provenance, measurements, honest non-claims, lane/convention notes, exact manual commands, and the decisive/skip policy.

What this smoke proves

• A managed Chrome launch through the shared headless wrapper starts, reaches Ready, and records runtime browser/protocol identity.
• At least 30 non-empty JPEG frames are captured with strict per-target CaptureOrdinal continuity, non-decreasing observed/session clocks, and JPEG dimensions matching metadata.image().
• A capacity-one, blocked-sink session declares IngestionQueueSaturated honestly with a positive missing-frame estimate and drains accepted work before a bounded managed stop.
• No managed Chrome process or temporary profile reference outlives its session on Linux (/proc/*/cmdline) and macOS (ps -ax -o pid= -o command=).
• The committed macOS evidence covers the default-DPI configuration. A high-DPI configuration was attempted, but it produced no passing artifact; the decisive observed device-scale threshold remains >= 1.5.

Honest non-claims

Every evidence document carries a non_claims array repeating these limits:

• No transport requalification. The cdpkit 0.4.0 selection, 60 s / 1 000-frame sustained saturation, deterministic wire assertions, and ack p99/max thresholds are owned by docs/evidence/cdp-transport/v2/.
• No host-speed percentile threshold. Ack/cadence histograms are diagnostics with internal consistency checks only (samples ≥ N, percentiles monotonic); no p99 ≤ X ms claim is made.
• No product-thesis capture-probability threshold. The 100 ms / 95 % and 50 ms / 80 % capture envelope and the 25 pp agent-improvement bar belong to the evaluation epic.
• No duration sweep, visual-defect corpus, artifact comparison, or storage validation. Those are owned by the evaluation epic.
• No claim that Chrome's acknowledgement token changes, orders frames, or detects skipped browser frames.

Measurements

Each session records:

• frame_count, source_time_samples, image_dimensions, viewport, device_scale_factor
• capture_ordinal_range.min and .max
• observed_clock_span_nanos, session_clock_span_nanos
• ack_latency_nanos and frame_cadence_nanos as {samples, p50, p95, p99, max}
• Declared gaps as {reason, count}
• visibility_events count (usually zero)

The provenance block records git revision, Rust toolchain, cdpkit version, discovered installation, runtime BrowserVersion, the forced-DPI wrapper variant, and a snapshot of the actual CaptureConfig used (fidelity uses CaptureConfig::default() verbatim; loss-reporting overrides only queue_capacity to 1).

Lane/convention notes

• Evidence is written by the test to $KROMETRAIL_SMOKE_EVIDENCE_DIR (default: a unique temp path). The test never mutates the source tree during cargo test; the operator commits the produced JSON.
• The smoke is opt-in via KROMETRAIL_REAL_CHROME_TESTS=1.
• ChromeWrapperVariant::DefaultDpi forces --force-device-scale-factor=1; ChromeWrapperVariant::HighDpi adds --high-dpi-support=1 --force-device-scale-factor=2.
• Linux Chrome and Linux Chromium are selected explicitly by BrowserProduct, not by first-discovered order.

Manual evidence procedure

Linux lane (Chrome required; Chromium optional)

KROMETRAIL_REAL_CHROME_TESTS=1 \
KROMETRAIL_SMOKE_EVIDENCE_DIR=docs/evidence/cross-platform-smoke/v1 \
cargo test -p krometrail-cdp --test cross_platform_smoke --locked -- --nocapture --include-ignored

This writes linux-chrome.json always and linux-chromium.json only when a BrowserProduct::Chromium installation is discovered. If Chromium is absent, the run prints not_installed and continues.

macOS lane (Chrome required)

KROMETRAIL_REAL_CHROME_TESTS=1 \
KROMETRAIL_SMOKE_EVIDENCE_DIR=docs/evidence/cross-platform-smoke/v1 \
cargo test -p krometrail-cdp --test cross_platform_smoke --locked -- --nocapture --include-ignored

This writes macos-chrome-default-dpi.json and attempts the high-DPI configuration. The high-DPI document is emitted only when production metadata observes device scale >= 1.5; the current attempted run emitted no macos-chrome-high-dpi.json.

Decisive/skip policy

• Decisive configurations for a complete smoke result remain linux-chrome, macos-chrome-default-dpi, and macos-chrome-high-dpi; the high-DPI threshold is not relaxed.
• Best-effort configuration (honest skip permitted): linux-chromium.
• The current high-DPI artifact is absent because the attempted run did not pass the threshold. This documentation records the non-pass honestly and does not block the local-tool release on absent platform evidence.
• Wrong platform: the test skips with a printed reason on any target_os outside {linux, macos}.
• Skip, not fail: when KROMETRAIL_REAL_CHROME_TESTS != 1 or no decisive installation is discovered, the test skips cleanly with the exact reason printed to --nocapture output.

Evidence that exists vs. honestly absent

• linux-chrome.json is present and passed both fidelity and capacity-one loss-reporting sessions against Chrome 149.0.7827.155 on Linux x86_64.
• macos-chrome-default-dpi.json is present and passed both sessions against Chrome 149.0.7827.201 on hosted macOS 14 arm64. Its process/profile cleanup result used the real ps-based reference scan.
• macos-chrome-high-dpi.json is absent because the attempted high-DPI runs did not pass. Hosted runs 29288505121 and 29288634536 both observed deviceScaleFactor == 1.0 under --high-dpi-support=1 --force-device-scale-factor=2; the production metadata observed deviceScaleFactor == 1.0, so the decisive >= 1.5 assertion failed and the harness emitted no passing document. This is an attempted configuration, not high-DPI pass evidence. The threshold remains unchanged: a future lane must either demonstrate a Chrome configuration that exposes high DPI through production capture metadata or redesign that capture boundary before this artifact can be added.
• linux-chromium.json is honestly absent because no BrowserProduct::Chromium installation was available on the Linux lane; this configuration remains best-effort.
• The temporary publication used branch tmp/cross-platform-smoke-macos-595f079, workflow CDP transport macOS gate, workflow patch commit a99d18e1c0bb8e07fed80c278cc05e59494bb21a, and evidence commit 4dfc78b2a6e91c3404fc58c3c8b98c5b6d662fdc. Run 29288505121 first exposed the high-DPI failure; run 29288634536 retained and uploaded the valid default-DPI document before the same decisive high-DPI failure.
• Real visibility transitions and Page.screencastVisibilityChanged events are observed and counted, not assumed.