Per-chip ARQ capability matrix: ack_responder/retry caps flags, 8821AU responder root-cause, Kestrel retry H2C, 8814A carve-out audit

[josephnef]

Context

Hardware ARQ is a per-chip capability and the current coverage is uneven, partly measured and partly folklore-adjacent:

• 8821AU as responder: bench-measured broken — an armed 8821AU never closed the loop (TX retries stayed pinned) while the 8814AU ACKs ~100% with the identical recipe.
• 8814A die: retry knob inert — the vendor DATA_RETRY_LIMIT=0 carve-out is preserved; nobody knows if it hides a fw bug or is cargo cult.
• Kestrel: retry is fw-level — DEVOURER_TX_RETRY_LIMIT is a no-op there.
• Untested as responder: Kestrel 8852B/C, PCIe 8821CE, 8812CU/8822EU (the 8812EU is now heavily validated via ARQ end-to-end delivery bench: three-ledger harness; count spsc-fat pool-exhaustion drops #359).

Callers currently discover all of this in the field.

Steps

1. [measure] Responder matrix. tests/ack_responder_check.sh across every owned die as responder against a J3 reporter TX; reconfirm the 8821AU failure; first data for Kestrel + PCIe. Output: one table.
2. [build] Caps flags. AdapterCaps: ack_responder_ok + tx_retry_limit_ok, bench-derived truth tables doc-commented at the declaration (the AdapterCaps convention — the header is the source of truth), resolved at construction, carried in the adapter.caps event.
3. [research] 8821AU responder root-cause. Per the match-chip-dispatched-source rule: diff the vendor 8821A-die branch for MSR/MACID/RCR handling vs the 8812 path before touching registers; fix if a recipe variant emerges, else document the exclusion in the caps table.
4. [build] Kestrel retry knob. Port the mac_ax retry-limit H2C so DEVOURER_TX_RETRY_LIMIT means the same thing on AX; validate with a Kestrel TX arm + responder.
5. [research] 8814A carve-out audit. Source archaeology for the vendor's DATA_RETRY_LIMIT=0 rationale, then a cautious bench enable (the die has known TX-wedge modes — watch rc=-7 patterns). Acceptable outcome: keep the carve-out, but documented as measured rather than inherited.

Acceptance

Matrix table in docs/scheduled-mac.md; caps flags shipped; 8821AU and 8814A entries backed by either a fix or a documented, bench-confirmed exclusion.

Refs: #359, docs/scheduled-mac.md, src/AdapterCaps.h conventions.