Capability proxy: kitty flag state + remaining static-query coverage

[chaodu-agent]

Deferred from the #16 / capability-proxy follow-up review (fable). Non-blocking refinements:

• Kitty keyboard is stateful, answered as static. The proxy always answers CSI ? u with ?0u, but CSI > flags u (push) / CSI < u (pop) / CSI = flags ; mode u (set) pass through and change the client emulators key encoding. An app that pushes then queries is told the push failed. The proxy sees the push/pop/set go by and could track the flag stack, or the behaviour should be documented.
• Static-query coverage gaps (pass through today, so the client answers and — when the filter is on — TermFilter strips the reply, i.e. a dead-end): OSC 12 (cursor colour), OSC 4 (palette), tertiary DA (CSI = c), DSR ?-family (?25n/?26n/…), multi-param OSC 10;?;? (SwiftTerm answers fg+bg), and zero-padded params (CSI 00c/05n — SwiftTerm parses numerically, the proxy compares bytes).
• OSC terminator: proxy mirrors the query terminator (BEL→BEL); SwiftTerm always sends ST. Harmless, arguably better (xterm-like), but not byte-exact as the module doc claims — normalise or adjust the doc.
• CPR still enters the ring buffer (not consumed at source), so a ?since= replay can re-deliver a CPR query; the client must answer only live CPR. Endgame is answering CPR at the source via an embedded VT state machine (libghostty-vt), which also removes it from replay. See Terminal capability proxy: the runtime should answer static queries itself #15.