feat: <Esc> dismisses, and it counts

[teocns]

Closes the discoverability hole around dismissal, and ports the two rejection mechanisms behind Cursor's Escape.

The problem

<C-]> was bound and worked. It appeared in one un-highlighted README table row, never beside a suggestion, so the reasonable conclusion was that dismissal did not exist.

The fix is not a new keymap

<Esc> already dismissed — InsertLeave clears the suggestion — it just filed nothing, so the same edit returned the moment you re-entered insert. Now it records the rejection, and the hint names the key that was working all along:

⟪neocursor · <Tab> accept · <Esc> dismiss⟫

No insert-mode <Esc> mapping is installed. In VS Code Escape is a free key; in neovim it is the mode transition. A conditional Esc would make macros non-deterministic (whether a ghost is showing depends on a debounced network reply), clobber im-select and snippet mappings, and has no clean fallthrough — feedkeys with "n" skips the mapping the user wanted, "m" re-enters ours. Advertising behaviour that already exists costs none of that.

<C-]> stays as the dismiss-without-leaving-insert variant. It is the key copilot.vim, copilot.lua and avante.nvim all use.

This is what Cursor does under a vim layer

handleKeyDownForCppKeys is a raw document.addEventListener("keydown", …), not a registered keybinding, and the suggestion branch never calls preventDefault — only the auto-import branch does. So in Cursor + vscode-neovim, Escape dismisses the suggestion and propagates to neovim, which leaves insert. Both happen; their tiering collapses to one tier for vim users by accident.

Tiering, where it can be kept

Cursor's first Escape files the suggestion as rejected and deliberately keeps the jump target alive (maybeShowHintLineWidget fires immediately after); only a second press, with nothing showing, reaches clearCursorPrediction. reject_suggestion() / reject_prediction() split accordingly and <C-]> walks them in order.

Collapsing the two — which the first draft did — mutes jump targets the user never said no to for the 30s pred_rejects TTL, after two ordinary trips through normal mode. flow_spec round 4 caught it.

Churn gate

Ports hasRejectedTooManySuggestions, which the per-suggestion ledger cannot express:

	catches
HEURISTIC_SUGGESTING_RECENTLY_REJECTED_EDIT	the model repeating itself
this	the model being wrong in a new way every time

Past the budget the passive triggers stop firing — entering insert, moving to another line — while typing still asks. That split is Cursor's: their content-change and linter-error paths never consult the gate. Accepting anything refills it, as does leaving the buffer (onDidBlurEditorText; merely leaving insert is not the analogue, that happens constantly here).

One deliberate divergence. Cursor increments on every clearSuggestions(); this counts dismissals. Their suggestion survives typing via isOnShortestEditPath, ours only does for inline ghosts — a diff is dropped and refetched per keystroke, so counting clears would mute after ~20 characters rather than ~20 ignored suggestions. Documented at the call site.

maxNumberOfClearedSuggestionsSinceLastAccept now arrives with the rest of CppConfig; the live backend returns 20, which the hardcoded default independently matched. :NeocursorLog shows the tally and flips to MUTED at the threshold.

Testing

Full CI suite green — flow_spec (both hint modes), hints_spec, paths_spec, handshake_spec, docs_links. No new spec: the behaviour is exercised through existing rounds, and round 4 proved it has teeth by failing on the first draft.