Follow-ups left open by TxMode::no_agg / NOAGG (#459)

[josephnef]

Left open by #459 (TxMode::no_agg / /NOAGG, merged as 87dde96). None blocks the feature as shipped; each is a measurement or a port the PR explicitly did not claim.

• Cost of occasional flagged frames. Only the every-other-frame worst case was measured (−47 % heard on the 8812EU, −60 % on the 8812CU). The realistic shape, one control frame every N aggregated video frames, is unmeasured. tests/tx_no_agg_onair.sh needs a duty knob (flag every Nth frame) and a goodput reading, not a heard-fps one.
• Mixed-width fold. The A-MPDU folds bandwidth as well as rate (a 20 MHz frame behind a 40 MHz opener airs at 40). The harness scores rate only. Add a width arm once the witness's rx.seq carries the decoded bandwidth.
• Port to Jaguar1 / Jaguar2 / MT7612U. tx_no_agg_ok is false there by design. Jaguar2 shares the halmac descriptor (SET_TX_DESC_BK dword2[16]) so it is likely the same two writes; Jaguar1's 8812A descriptor and the MT7612U TXWI need their own look. Each needs an on-air arm, not a code-reading.
• Cross-rate fold on the 8822C is deeper than on the 8822E (24 % vs 40 % of MCS0 frames kept their rate unflagged, same script, same knobs). Unexplained. Candidates: USB high-speed vs SuperSpeed feed depth, or a die difference in how many MPDUs the MAC waits to accumulate. One unit per die so far.
• 8812EU on the OpenIPC bench drops off the USB bus about 1 s after the TX bring-up reset (dmesg disconnect + re-enumerate; doctor HEALTHY on EFUSE/fw/RX). It ran a 60 s flood on 2026-09-26, so the bench changed. Rig, not code, but it is why the EU arm was not repeated here.