fix(openssh): close residual regress compatibility gaps

Part of #275. Follow-up to #313.

## Context

The final OpenSSH `V_10_3_P1` compatibility runs for #313 reached the epic gate on both CI platforms:

- macOS: 60 passed out of 66 eligible tests, 6 failed, 7 environmental, 6 skipped
- Ubuntu: 61 passed out of 69 eligible tests, 8 failed, 4 environmental, 6 skipped

This issue keeps every remaining candidate-only failure explicitly tracked after the epic closes. It does not reclassify environmental failures or reduce the committed score floor.

## Remaining candidate failures

| Test | macOS | Ubuntu | First observed gap |
|---|---|---|---|
| `agent-restrict` | fail | fail | candidate key generation path |
| `dynamic-forward` | fail | fail | dynamic forwarding startup and SOCKS coverage |
| `hostkey-agent` | fail | fail | Ed25519 host-key agent behavior |
| `multiplex` | fail | fail | unsupported `-O conninfo` and remaining control operations |
| `percent` | fail | fail | percent expansion in match/exec handling |
| `proxyjump` | fail | fail | basic parsing and multi-hop behavior |
| `rekey` | environmental | fail | Ubuntu candidate timeout |
| `scp3` | environmental | fail | Ubuntu remote-to-remote copy behavior |

The uploaded result tables classify `channel-timeout`, `connection-timeout`, `forwarding`, and `sftp-badcmds` as environmental on both platforms. macOS additionally classifies `forward-control`, `rekey`, and `scp3` as environmental because the reference client fails in the same environment. The six upstream skips remain unchanged.

## Scope

- reproduce each candidate-only failure against the pinned harness
- fix compatibility gaps without weakening fixture validation or score floors
- split independent implementation areas into linked issues when a focused fix is safer
- retain reference-client evidence for platform-specific environmental classifications

## Acceptance criteria

- [ ] Every listed candidate-only failure either passes on both supported CI platforms or has a narrower linked issue with reproduction evidence.
- [ ] The committed eligible-result floors are not lowered to hide a regression.
- [ ] macOS and Ubuntu result artifacts identify all remaining fail, environmental, and skip verdicts.

Start by reproducing the listed candidate-only failures against the pinned harness on macOS and Ubuntu, using the result tables and reference-client evidence to distinguish environmental failures. Track agent-restrict, dynamic-forward, hostkey-agent, multiplex, percent, proxyjump, rekey, and scp3 separately, and split independent areas into linked issues when appropriate. Done means each candidate failure passes on both CI platforms or has a narrower linked issue with evidence, without lowering score floors.

Written by the indexing model from the issue text.