feat(hub): declarative pane-state detection + generic pane-input hardening (P1, Q1)

[physercoe]

Pane-state-manifests plan, first wave — P1 (the core evaluator, pure library) and Q1 (the smallest independent).

P1 — hub/internal/panestate

idle.go is one regex plus a 90-second stall (idle.go:25,49,55), and it is the only state signal for every engine without a structured M4 adapter. codex parked on "Allow command?" raises nothing today.

This makes detection declarative: an evaluator for herdr's manifest schema plus its 19 per-agent TOMLs, vendored byte-exact at 6f311498 (Apache-2.0, NOTICE updated) and pinned by git blob SHA so a re-vendor diffs cleanly. Rules are data — the ADR-010 frame-profile shape — so a new engine is a TOML, not a Go adapter. No wiring (D-8, teeth before wiring).

Everything of ours is in manifests/overlay/, including the family mapping. claude-code, codex, antigravity, gemini-cli are mapped; kimi-code-ts deliberately is not — upstream detects a CLI it calls kimi and nobody has confirmed that is our compiled-TypeScript one. Screen rules are more version-fragile than a resume flag, so the bar is a real capture, not a name match. An unmapped family gets no evaluation: another engine's rules would produce confident wrong attention, which is worse than today's silence.

Two porting calls that needed judgement

1. Rust-regex and RE2 are not the same dialect. 9 of the 58 vendored patterns don't compile in Go — 5 use \uXXXX escapes, 4 use \p{Alphabetic}, a Unicode binary property RE2 lacks. Editing the vendored files is forbidden (D-1) and dropping the rules would be a silent hole, so regex_translate.go rewrites at compile time and records each translation, flagging the inexact one:

from	to	exact?
\uXXXX / \u{XXXX}	\x{XXXX}	yes
\p{Alphabetic}	[\p{L}\p{Nl}]	no — drops Other_Alphabetic

Every vendored use is \p{Alphabetic}+\w*ing\b after a spinner glyph, so the dropped set is unreachable there — but that's a claim about the manifests as they are today, so it's labelled rather than assumed, and a test fails if the inexact flag disappears.

2. An unimplemented region is refused, not emptied. Upstream resolves an unknown region to "", which turns a typo or a newer-schema region into a rule that silently never fires — the exact failure class this plan exists to remove. Only the 8 region kinds the corpus uses are implemented; the rest fail validation by name. That's the plan's risk 3 working as intended. Same reasoning for unknown TOML keys: an unread key is a rule that never fires, so it's an error.

Q1 — generic pane input

PaneDriver serves every engine without an adapter, and it sent a multi-line body with one send-keys -l — so the pane submitted at the first newline and the rest arrived as separate turns. The per-engine adapters were fixed for exactly this; the fallback path never was. It now takes the same named-buffer paste with a single explicit Enter.

All four paste paths gained -p. Verified against tmux(1) rather than the discussion's paraphrase: "paste bracket control codes are inserted around the buffer if the application has requested bracketed paste mode" — so a TUI that asked for it stops line-editing the body, and one that didn't is unaffected. The generic path also waits 300 ms before Enter so the TUI has ingested the paste. Short single-line input keeps its original two-call shape, so no existing interaction changed.

One guard worth a look: a harness that stubs SendKeys but not Tmux gets a refusal, never a fall-through to a real tmux exec. On a developer box that server is the one the human is sitting in.

Verification

The corpus is 14 of upstream's own test screens with upstream's own expected answers (claude, codex, devin), so it's a cross-implementation parity check rather than a self-consistency one. Building it turned up an extraction bug worth naming: reading the nearest following assertion attributed one screen's expected rule id to the next screen in the same test function — the window is now bounded to the current call.

Coverage is narrower than the plan's line implies, and I'd rather say so than imply otherwise. The other 16 manifests get parse + validate + compile + empty-screen fallback only. Per-agent blocked/working screens need real captures; inventing them from the rules under test would prove nothing. That's device-verify debt the plan already books.

Mutation-checked:

• tie-break flipped to last-wins → named tie-break test fails
• contains made case-sensitive → 10 tests fail, including corpus cases
• -p dropped → flag test fails
• multi-line path reverted to one send-keys → block test fails

Full go test ./... exit 0; 11 runnable CI lints pass; gofmt/vet clean.

New dependency

github.com/BurntSushi/toml v1.4.0 — BSD-2, zero transitive deps, go 1.18 so it clears this module's 1.23 pin (the same constraint that keeps parquet-go at v0.25.0). D-1's byte-exact vendoring requires parsing upstream's format; I picked the smaller of the two candidates.

🤖 Generated with Claude Code