no dash punctuation left in the prose, and copilot's file is gofmt clean

Two housekeeping passes the repo had already committed to and had not
finished.

gofmt: internal/provider/copilot/responses.go arrived unformatted with #15.

Dashes: the em dash sweep of ecce583 cleared the docs a reader opens first
and left the deeper ones, the Go comments, and everything written since. An
em dash becomes a comma or a colon, a numeric range becomes a hyphen, and
nothing else moves.

What is deliberately NOT swept, because it is not punctuation: the pager
footer paints `N-M/T` with an EN DASH, `transcript.go` formats it that way,
`paint-jogdiff.sh` and `paintpane.sh` grep for that glyph, and the docs that
quote a captured footer have to quote it exactly. Changing the footer is a
UI change with test and script coupling, not a typographic one, so every
line depicting it is left alone: aria.go, transcript.go, the three paint
scripts, paint-repro.md and tmux-testing.md. Recorded tapes and golden files
are evidence and were never candidates.

Author: gluck

README.md

@@ -8,7 +8,7 @@
 
 *Largo al factotum della calcolatore.*
 
-**[figar.org](https://figar.org)** — the site. Docs coming soon.
+**[figar.org](https://figar.org)** is the site. Docs coming soon.
 
 A coding agent with few opinions, firmly held. Supports multi-agent, context forking, fully persistent structured metadata, and terminal-native semantics out of box.  One binary: CLI, supervisor daemon, agent runtime. JSON-RPC over unix sockets.  Tutto pulito, bravo bravissimo.
 

internal/cli/docs/completion.md

@@ -200,7 +200,7 @@ figaro x -- list files in @cwd!  # @cwd! expanded before instruction
 - Hidden files are never suggested by `listCWD`. Fish covers them
   via native file completion; bash/zsh do not.
 - All RPC fetches (aria ids, live snapshot keys, snapshot for
-  `@`-expansion) bound at 300–500ms. A sluggish daemon can make tab
+  `@`-expansion) bound at 300-500ms. A sluggish daemon can make tab
   feel laggy or expansion fall through to literal, but neither
   blocks long.
 - Upgrading figaro does not regenerate the on-disk completion script.

internal/cli/keymap_equiv_test.go

@@ -59,7 +59,7 @@ func oracleTranscript() *transcript {
 // It is relative because the absolute offset is not what this oracle is for.
 // Following the tail pins the offset at `total_lines - body`, so any change to
 // how a message is drawn moves every recorded number while changing no
-// binding at all — which has now happened often enough to be the main cost of
+// binding at all, which has now happened often enough to be the main cost of
 // owning the table. A delta is a statement about the KEY ("u scrolls a page
 // up"), and a page is a function of the pane, not of the conversation.
 //
@@ -75,7 +75,7 @@ func offToken(tr *transcript, base int) string {
 		return "off:top"
 	case tr.follow && off == maxOff:
 		// FOLLOWING: the offset is not a position the key chose, it is a
-		// consequence of being pinned to the tail — entering the transcript,
+		// consequence of being pinned to the tail, entering the transcript,
 		// toggling verbose and opening a panel all re-pin it. Recorded by name
 		// so the number, which is content length, stays out of the table.
 		//
@@ -101,7 +101,7 @@ func oracleSignature(tr *transcript, base int) string {
 		len(tr.expanded), tr.inJump, tr.jumpQuery)
 }
 
-// inertOffset is where the offset sits after a key this state does not bind —
+// inertOffset is where the offset sits after a key this state does not bind,
 // the reference every cell in the row is measured against. The probe key is
 // chosen from the row itself, so a state that later binds 0x01 cannot silently
 // start measuring against something that moves.

internal/cli/keymap_regen_test.go

@@ -17,8 +17,8 @@ import (
 //	ORACLE_REGEN=1 go test ./internal/cli/ -run TestOracleRegen
 //
 // Both regenerators press their keys through the same sweep the verifier uses
-// (sweepPager, sweepInput) — a second copy of the sweep could record cells the
-// verifier never reads — so all that is left here is the formatting.
+// (sweepPager, sweepInput), a second copy of the sweep could record cells the
+// verifier never reads, so all that is left here is the formatting.
 
 // oracleRows renders one row of a frozen table: the inert signature, then
 // every cell that differs from it, sorted so a regeneration diffs cleanly.

internal/cli/speaker_header_test.go

@@ -54,7 +54,7 @@ func TestSpeakerHeader_FirstSliceIsCorrect(t *testing.T) {
 
 // A CONTINUATION slice of the same turn carries no question, so it must not
 // announce the speaker again: the header belongs to the inquiry seam, and a
-// continuation has no seam. This is the reported bug — a "< figaro" printed
+// continuation has no seam. This is the reported bug: a "< figaro" printed
 // between two tool nodes of one turn, at the page boundary that cut it.
 func TestSpeakerHeader_ContinuationCarriesNone(t *testing.T) {
 	m := aria.Message{

internal/cli/speaker_header_wire_test.go

@@ -19,14 +19,14 @@ import (
 //
 // This closes it from the other end:
 //
-//	aria.PaginateBefore  — the real byte-budget walk, at the real 64 KiB
+//	aria.PaginateBefore: the real byte-budget walk, at the real 64 KiB
 //	  -> assemble         sets ClippedHead when the window opens mid-turn
 //	  -> committedMessages drops the inquiry for that part
 //	  -> Composer.Message  must therefore draw no speaker header
 //
 // The turn sizes are not invented either. Measured over four of this store's
-// own arias, turns over the 64 KiB budget are ordinary — 4 of 24, 3 of 18, 9 of
-// 15 — with the largest at 343 KB, five times the budget. A page walking back
+// own arias, turns over the 64 KiB budget are ordinary, 4 of 24, 3 of 18, 9 of
+// 15, with the largest at 343 KB, five times the budget. A page walking back
 // through any of them opens mid-turn by arithmetic, not by luck.
 const productionPageBudget = 65536 // config.defaultPageBudget
 

internal/cli/stream.go

@@ -795,7 +795,7 @@ func (in *interactiveInput) pageTranscriptSearch(ctx context.Context, cancel con
 // part carries our own text; 4 dials, warm aria):
 //
 //	Qua returned          +0.13 … +0.15 ms
-//	our inquiry frame     +0.19 … +0.34 ms   (0.04–0.19 ms after Qua returned)
+//	our inquiry frame     +0.19 … +0.34 ms   (0.04-0.19 ms after Qua returned)
 //
 // which agrees with W1B's BenchmarkPromptBroadcastGap: the append→broadcast
 // window inside the daemon is ~0.7µs on a 5,000-message aria and flat in aria

internal/cli/transcript_jump.go

@@ -54,7 +54,7 @@ import (
 //
 // 24 is chosen against measurements this repo already has, not by feel. One
 // fetch is a single ReadBefore round trip to the local daemon over a unix
-// socket: 0.1–1.2 ms measured on the long-aria benchmarks, and returns
+// socket: 0.1-1.2 ms measured on the long-aria benchmarks, and returns
 // pageMessages() messages, which is between transcriptMinPageSize and
 // transcriptPageSize. So 24 fetches cover several hundred to a few thousand
 // messages, comfortably past the largest real aria measured on this project

internal/cli/transcript_jump_deep_test.go

@@ -14,7 +14,7 @@ import (
 // A store that HOLDS history and hands it over a page at a time.
 //
 // Every other jump fixture in this package answers a page request with the
-// empty page — the floor — so the walk terminates on the first fetch and the
+// empty page (the floor), so the walk terminates on the first fetch and the
 // only paths exercised are "already here" and "does not exist". The reported
 // failure is neither: `:5` on a large aria showed "jumping to turn 5…", pulled
 // one round of history, and then stopped, silently, at the tail. That needs a

internal/cli/transcript_resize_paint_test.go

@@ -20,7 +20,7 @@ import (
 //	 area. It will typically be fixed upon return."
 //
 // Reproduced in a real tmux pane before this test was written (see
-// skills/figaro/contributing/paint-repro.md §8): at 100x40, scrolled to rows 219–240 of 1058, a width
+// skills/figaro/contributing/paint-repro.md §8): at 100x40, scrolled to rows 219-240 of 1058, a width
 // change to 72 columns left FIVE rows holding text from other lines, and a jog
 // away and back cleared them.
 //