Packaging era split: 25.12 SDK cannot emit ipk — two pinned SDKs (format-selected), or drop ipk #708

The 24.x/25.x packaging split — analysis and proposal (from the rc1 build driver failures)

Observed: the ipk-lane driver failed on the 25.12.0 SDK with "no .ipk found" — because OpenWrt 25.12 is the apk-native era: its SDK emits .apk regardless of the CONFIG_USE_APK dance (scripts/build-sdk-package.sh appends CONFIG_USE_APK=y only for the apk branch; 25.12's default is already apk, so the ipk lane searches for an extension that era never produces). The rc1 build's apk lane succeeded (tollgate-wrt-0.6.0_rc1-r0.apk staged); the ipk lane cannot succeed on this SDK by construction.

Why ipk can't just be dropped: fork #93/#110's evidence — the installed base installs via opkg on feed-less 24.10-era systems (and orphan-removal semantics that #707 fail-louds). The ipk output is the format those devices consume.

Proposal — two pinned SDKs, format-selected:

1. packaging/build-inputs.json: extend openwrt_sdk from a single release: 25.12.0 to per-format pins — apk: {release: 25.12.0, …} and ipk: {release: 24.10.x, …} — each with its own digest-pinned targets. The scaffolding already exists: go_per_release maps 24.10's toolchain today, and sdk-go-version.sh audits per release; the .go.version pin follows the apk (release) line as the shipped-binary truth, with the ipk SDK's older Go checked as ≥ directive floor (and if 24.10's feed Go < 1.26: the ipk lane pins GOTOOLCHAIN from the manifest tarball — the same single-source-of-truth rule ci(toolchain): two paths, one truth — manifest-driven parity gates for the SDK and shortcut build paths #687 established).
2. build-sdk-package.sh: PACKAGE_FORMAT selects its SDK; the driver's PACKAGE_EXTENSION check then matches what its era actually emits.
3. CI matrix: arch × format → the right SDK per row.

Owner decision embedded: confirm ipk stays a supported output (recommend yes — the installed base needs it). If instead the answer is "25.12/apk only", the ipk matrix rows delete and this issue closes with the era note.

(From the orchestrator session; the release lane is fighting this exact mismatch in the rc1 driver right now — implementation can ride its current build work.)