Bump to 0.7.31 — frame-tracker ownership fix + honest platform docs

[GenericJam]

First run of the release gate that landed in 0.7.30. Reviewing 0.7.29..master caught three things in the MOB-103 work that 0.7.30 already shipped.

1. iOS frame tracker lost its ownership token on a refused write (real bug)

record assigned box.seq unconditionally, so a generation-refused write (returns 0) overwrote a valid seq with 0 — and mob_unregister_frame ignores seq 0, permanently disabling that tracker's own cleanup.

Failure: push from a screen tagging :id "save" to one whose "save" is a lazy row below the fold. The outgoing screen's .move writes are refused by the generation gate, box.seq becomes 0, its .onDisappear is a no-op, the id stays in the tree so the purge keeps it — and Mob.Test.tap_id taps the old screen's coordinates.

Keeping the last successful seq is strictly safe: the compare-and-delete still refuses to delete whenever an incoming tracker has claimed the id since. The .opacity (reset) transition doesn't hit this, which is the tell — it doesn't move the outgoing frame, so nothing gets refused.

2. element_frames/1 docs implied cross-platform parity that doesn't exist

The drop-when-not-laid-out behaviour is iOS only. Android clears frames wholesale on a navigation transition and never per element (MobBridge.setRootJson), so a scrolled-away lazy_list row keeps its frame and tap_id happily taps it. Verified in the template source. Docs now say so plainly instead of implying parity; Android parity wants its own issue.

3. 0.7.30 published with empty release notes

CHANGELOG's top section was still ## [Unreleased] when 0.7.30 was tagged, so release.yml's ## [<version>] extraction found nothing. Retitled to [0.7.30] so the notes exist for anyone reading back. Exactly the version/release-seam class the gate exists to catch.

Also

Wires make -C test/native run into the macOS CI job. That harness is the only automated guard on the registry's ownership/generation/purge rules, and it reproduces the algorithm rather than linking it — running it is what keeps the copy honest. It was committed in 0.7.30 but nothing ran it.

Preflight

mix format --check-formatted, mix credo --strict, mix compile --warnings-as-errors, 1111 tests, clang-format, and the native harness (20/20) all clean.

🤖 Generated with Claude Code

[GenericJam]

mob_send_dismiss device-verified (physical Moto G power, 2021)

Verified through the full JNI chain rather than just statically:

onDismissRequest → MobBridge.nativeSendDismiss → Java_com_example_sheetprobe_MobBridge_nativeSendDismiss → mob_send_dismiss → sendEvent(handle, "dismiss") → {:dismiss, tag} delivered to the screen's handle_info/2.

The probe screen matched both {:dismiss, tag} and {:tap, tag} so a wrong shape would be logged rather than kill the screen. Result across two open/dismiss cycles: log: [:dismiss, :dismiss], the :tap clause never fired, screen alive throughout, sheet re-presented correctly after each dismissal.

Paired with GenericJam/mob_new#38, which supplies the nativeSendDismiss extern and JNI thunk.

Release ordering: mob_new's CI clones GenericJam/mob at master with no ref pinned, so merging this PR unblocks mob_new's CI immediately — it currently fails with call to undeclared function 'mob_send_dismiss', which is the lockstep working as intended.