[iOS] Investigation umbrella: consolidated mobile connectivity day (DCUtR verdict, NWPathMonitor premise, push-budget reality, NSE wake paths)

[pvg13]

Objective (M1 / P1)

One hardware day that resolves the open iOS connectivity cluster with device evidence and produces the iOS half of M1's mobile picture. Deliverable: a report in docs/research/ with verdicts that close or convert the member issues.

Member issues (all assigned to M1)

• [iOS] Validate DCUtR hole-punching on cellular; confirm relay-only sessions sync #77 — validate DCUtR hole-punching on cellular; confirm relay-only sessions sync. Extend with the byte-split diagnostics so the verdict is a ratio, not just success/failure.
• [iOS] Replace 3s interface-watch polling with NWPathMonitor for sub-second re-listen #74 — NWPathMonitor vs the 3s interface-watch polling (gated on [iOS] Verify the if_watch IfWatcher hang premise on-device; simplify QUIC bind if unspecified is safe #73's verdict).
• [iOS] Verify the if_watch IfWatcher hang premise on-device; simplify QUIC bind if unspecified is safe #73 — verify the if_watch/IfWatcher hang premise on-device; simplify the QUIC bind if the unspecified bind is safe.
• Feeds [iOS] NSE: read app-written localized fallback JSON when push composition yields no content #92 scheduling (NSE fallback text) — same device day if time allows.

Additional questions beyond the member issues

1. Silent-push budget reality: measured delivery rate of budgeted silent pushes across a day of normal use (foreground/background/killed) — bounds how much wake-dependence can be engineered away vs must be tolerated.
2. Suspension vs live P2P: how long does a direct QUIC session survive backgrounding before iOS kills the socket, and does the resume path re-establish direct or degrade to relay circuits?
3. NSE wake paths: does an alert-push NSE invocation give enough runtime for a mailbox drain over the relay vs a direct catch-up, and which is actually attempted?

Method

Two iPhones + the existing device protocol docs (docs/ios-device-protocol-2026-07.md Sections C-E as the base), relay with fresh push budget, diagnostics snapshots per scenario.

Output

docs/research/ios-cellular-p2p-2026-08.md; #73/#74/#77 closed or converted to implementation issues; iOS rows in the M1 KPI table.

[pvg13]

Session prepared — one phone, one USB connection, one script (dev 5b04603; plan updated in docs/research/ios-cellular-p2p-2026-08.md).

tests-e2e/ios/run_device_session.sh (run on the Mac) now drives the whole day:

• S1 WiFi-baseline + cellular carrier-NAT classification (phone ↔ host writer, the topology the Android day proved decisive)
• S3 the [iOS] Verify the if_watch IfWatcher hang premise on-device; simplify QUIC bind if unspecified is safe #73 bind A/B — devicectl launch --environment-variables flips WAVESYNC_IOS_UNSPECIFIED_QUIC per phase, no Xcode scheme editing, stdio streamed to stamped captures
• S4 three [iOS] Replace 3s interface-watch polling with NWPathMonitor for sub-second re-listen #74 handoff trials with the laps computed from the captures (plus a free trial-0 from S1's flip)
• S6 suspension survival, S5 push-budget (skippable if the relay lacks APNs creds), N1 the [iOS] NSE: read app-written localized fallback JSON when push composition yields no content #92 NSE-fallback check (skippable without the NSE build)
• Output: a per-issue prefilled summary.md + raw logs; the only human actions are the ones iOS physically requires (Control Center flips, HOME, reading the banner), each prompted and timestamped.

Enabler that landed with it: an engine-side, env-gated m1-diag beacon (WAVESYNC_M1_DIAG=1) — the classification line is now emitted by the library itself, so the Mediterranea build gets it with a plain rebuild against this rev (the same rebuild picks up the #92 NSE template fix).

Checklist to schedule the sitting:

1. Mac with Xcode 15+ (xcrun devicectl), phone dev-mode on.
2. Dev-signed Mediterranea build against wavesyncdb ≥ 5b04603 (beacon + NSE fix); the NSE target for N1.
3. Test relay reachable from cellular, with APNs creds for S5/N1 (else SKIP_S5=1 SKIP_NSE=1).
4. Optional second iPhone for the one cell the script can't cover (cellular↔cellular).

Parser half is self-tested (--selftest, green); the devicectl half is Mac-only and unexercised until the sitting — budget the first 15 minutes for launch/signing friction.

[pvg13]

Mode change per discussion — observational is now the default: no orchestration script, no host writer, no prompts. Mediterranea is used exactly as a real user uses it (dev build), and the Mac reads the evidence in ONE connection at the end of the day:

1. Dev build sets WAVESYNC_M1_DIAG=1 itself at startup (icon launches carry no env) and exposes a hidden dev-settings toggle on the builder's ios_unspecified_quic_bind flag for the [iOS] Verify the if_watch IfWatcher hang premise on-device; simplify QUIC bind if unspecified is safe #73 A/B stints.
2. Live the scenarios: walk off WiFi (S1/S4), background the app (S6), household member adds items while it's locked ([iOS] NSE: read app-written localized fallback JSON when push composition yields no content #92 banner + S5 wakes).
3. End of day: sudo log collect --device-udid … --last 8h → flatten with the log show predicate → tests-e2e/ios/analyze_logarchive.sh (dev fb0b274) → per-issue summary.md (beacon timeline, [iOS] Verify the if_watch IfWatcher hang premise on-device; simplify QUIC bind if unspecified is safe #73 markers, [iOS] Replace 3s interface-watch polling with NWPathMonitor for sub-second re-listen #74 handoff laps off the device clock, bg_sync wake tables, NSE lines). Parsers self-tested.

Go/no-go smoke first (2 min): icon-launch the app, check Console.app for any wavesyncdb line — observational mode needs the Rust logs in os_log outside Xcode. If absent, add a tracing→os_log bridge to Mediterranea, or fall back to the orchestrated runner (5b04603), whose --console capture doesn't depend on os_log.

Data-safety rules from the previous comment stand: test household group, test relay for anything relay-configurable, no on-device reinstalls (Simulator owns the #92 cold-cache row).