P24-T03 — Performance and memory (budgets frozen, code-only slice)

Status: IN_PROGRESS on codex/phase-24-android-acceptance. Issue: #98 OPEN (not done). Entry: P24-T02 code slice landed (05b3fec); novice/device proofs deferred by user directive below. Binds B-BR-C01–C06 / BUC-C01–C05. No migration, no backend change, no iOS work, no production code changed in this slice.

User directive (2026-10-02, recorded in planning)

• No emulator or physical-device runs during the remaining phases. Validation on device/emulator happens manually only after ALL phases finish.
• Until then, run only code (JVM/unit) tests. This slice follows that rule: no connected*, no instrumented runs, no new device measurement claimed.

Frozen budgets (predeclared, from integrated evidence)

Area	Budget (frozen)	Source
Photo wire target / hard cap	150 KiB / 256 KiB	PortablePhoto + backend budgets.go (P15-T01, P21-T02)
Decoded frame	≤ 1600 px edge, ≤ 2 MP	PortablePhoto.sampleSizeForBounds (P15-T02)
Encode attempts / working-mem hypothesis	3 / 32 MiB	PortablePhoto (device proof of hypothesis deferred, see below)
Transient cache TTL	24 h	PortablePhoto.TRANSIENT_TTL_MILLIS
Discovery page size	1..100, default 20	DiscoveryQuery (P20-T01)
Cached-home regression reference	382 ms	P18 local exit (regression only, NOT a cold-start result)
Genuine cold-process start	TBD by manual validation	No value invented here

Code-level guards (already in tree, re-run as evidence, no device needed)

• PortablePhotoTest: budgets mirror forward contract, sample-size keeps 12 MP / 48 MP / panoramic frames inside 1600 px + 2 MP, wire-cap bounds, 24 h expiry.
• DiscoveryQueryTest + DiscoveryStationsRuleTest: pageSize 1..100 enforced, so list/scroll queries stay bounded by construction.
• SourceTimeBadgeTest + StationPriceRowA11yTest (P24-T02): presentation-only, no domain work in composables.

Explicitly deferred (not waived, not green)

• Genuine cold-process start, list scrolling, peak heap / encode / OCR timing, background replay and battery on the frozen matrix: NO measurement in this slice. Real low-end hardware row stays open. Claiming acceptance now would be false.
• These items join the post-all-phases manual validation batch owned by the user.

Validation (code tests only, per directive)

• ./gradlew :domain:test :application:test :data:testDebugUnitTest :app:testDebugUnitTest :app:assembleDebug --no-daemon (record counts in commit evidence). No connectedAndroidTest, no emulator boot.
• git diff --check clean; scoped secret review (no secrets/PII/GPS/photo content).