fix(asahi): add experimental HDMI reconnect recovery

[avillagran]

Summary

• Add an experimental, opt-in downstream Asahi HDMI reconnect kernel patch, pinned to AsahiLinux/linux commit ce9f2eba72c061a50b2d790450e90af3439d8c24.
• Track HDMI disconnect generations, request a real DRM modeset, and recover firmware power before programming the mode even when DRM's active state is unchanged. Preserve a newer disconnect and withhold swaps while recovery remains pending.
• Include an isolated build-only wrapper/package recipe, package validation, pinned-source RED/GREEN regression tests, extracted-C seam tests, and a dedicated CI job.
• Document curated hardware evidence, limitations, safe side-by-side installation and rollback procedures in docs/asahi-hdmi-reconnect.md.

Ready for maintainer review as an experimental, opt-in patch; not proposed for automatic deployment. No installer/update integration, package installation, GRUB changes, or reboots are performed by this PR. No raw journals, personal configuration, or prebuilt kernel package are included.

Validation

• Focused offline build/package-contract test: bash test/shell.d/asahi-hdmi-reconnect-test.sh.
• python3 test/asahi/reconnect.py: pinned baseline SHA-256 checks and patch applicability; baseline C seam produces exactly 11 expected assertion failures; patched seam reports zero failures; both source regression tests pass. The expected baseline RED is a regression check, not a full-suite pass.
• Separate Bash syntax checks for the build wrapper and PKGBUILD; bin/omarchy commands --check passes (465 commands).
• Repository-level git diff --check exits 2 on the embedded .patch file's required context-line prefixes (space before kernel tabs and space-only context lines). The patch is deliberately preserved byte-for-byte; the applied kernel-source diff passes its whitespace check in the focused regression runner.
• Included patch SHA-256 is dbf332ead80ad84d5fb681fe83fc92c89cb88e0d4e416142acbb80565aacedcf, identical to the hardware-tested functional patch.
• Prior hardware campaign: Apple MacBook Pro (16-inch, M2 Pro, 2023), candidate release 7.1.13-1-1-ARCH-hdmi-recover, 2560×1440 at 75 Hz. User confirmed initial image and three complete unplug/replug cycles, without compositor/session restart. Retained captures through the third reconnect had no flip_done timed out, commit wait timed out, Oops, BUG, or panic matches. An initial-boot swallowed-swap message also existed in baseline; this is not a claim of a completely clean boot log.
• Prior machine-specific build compiled the five affected translation units and completed Image/modules/DTBs and package validation. This is historical evidence for the exact functional patch, not a completed build with the newly generalized wrapper.
• Earlier local full-suite run failed; GitHub full-suite CI now passes: the prior ./test/all run exited 1. Its local log contains locale-sensitive benchmark arithmetic, jq parsing, and awk syntax diagnostics. Baseline comparison was interrupted, so these failures have not been established as pre-existing or unrelated. The local failure is retained for transparency; both GitHub suite jobs subsequently passed.
• All five GitHub checks passed: Syntax + shellcheck, test/all, tests/all, Asahi HDMI source + C seam, and install.sh on Arch Linux ARM.

Remaining validation and maintainer review items

• Complete a fresh native aarch64 build/package with this PR's generalized wrapper. Current wrapper coverage is syntax/contract validation, not end-to-end native build validation.
• Reboot back to stock, verify the graphical session and persistent stock boot default. Stock and candidate remain installed, but post-fix stock rollback bootability has not been exercised.
• Review/stress the gap between flush generation check and asynchronous firmware submission; the check does not serialize hotplug against submission.
• Exercise synthetic-vblank work coalescing, real event ownership/callback ordering and framebuffer lifetime under rapid hotplug. Synchronous seam mocks cannot establish these properties.
• Verify compositor liveness if disconnect after atomic check withholds a swap; recovery depends on a subsequent modeset-permitted commit.
• Expand monitor/mode, suspend/resume, session and HDMI hardware/firmware coverage. Compilation of both firmware translation units is not hardware coverage of both versions.

Review scope

The staged patch, build/package path, tests and documentation received a separate publication review. No new functional kernel edits were made after the hardware campaign. Concurrency/liveness and packaging end-to-end validation remain explicit limitations for maintainer review rather than being represented as solved by the unit seam. This candidate does not implement USB-C DisplayPort or M3 enablement.

Extended runtime follow-up

The user reported another reconnection hours later. A read-only check found the same recovery kernel running after approximately 11 hours 46 minutes of uptime, with zero flip_done timed out, commit wait timed out, DCP has crashed, or set_digital_out_mode timed out matches in the retained current-boot journal. The journal includes a later reconnect with power-on and successful mode completion at monotonic time 15244 seconds. HDMI was disconnected when inspected, so this is not an independent confirmation of a currently visible image. Additional swallowed-swap messages exist in the extended journal; no claim of an entirely error-free log is made. The post-fix stock rollback reboot and broader race/stress coverage remain untested.