docs: strip em dashes from plans/

Author: gluck

plans/api-coherence.md

@@ -1,4 +1,4 @@
-# API Coherence — the RPC surface, and the CLI's dependence on it
+# API Coherence: the RPC surface, and the CLI's dependence on it
 
 Opened 2026-08-21 by 44754e14 (forked from ede92072), at Gluck's request:
 "reorganize and consolidate and simplify the rpc api design and the cli's
@@ -7,7 +7,7 @@ form / aria is attended".
 
 ## 1. The prior art, found and dated
 
-**`plans/rpc-surface-tightening.md`, commit `fd83b09a`, 2026-05-05** — one of
+**`plans/rpc-surface-tightening.md`, commit `fd83b09a`, 2026-05-05**, one of
 four A/B/C/D refactor plans (inbox, rpc, dead-code, bootstrap). Its thesis:
 
 > Shrink the figaro JSON-RPC surface to its minimum viable shape: five
@@ -21,12 +21,12 @@ interface so dispatch left `protocol.go`.
 
 **It landed.** `figaro.qua` is the wire name today; `SetModel`/`SetLabel` are
 gone; `internal/figaro/server.go` owns the method table; the figaro-side
-`protocol.go` is **14 lines** — the plan hoped for ~60.
+`protocol.go` is **14 lines**, the plan hoped for ~60.
 
 **Then the plan was deleted** by `4b5f40b7` (2026-05-10, "migrate plans/ and
-docs/ to figaro-la/docs"). It is on no branch of this repo. Its sibling — the CLI architecture dossier added by `1cffcaaf`, deleted in
+docs/ to figaro-la/docs"). It is on no branch of this repo. Its sibling, the CLI architecture dossier added by `1cffcaaf`, deleted in
 the same migration, and named here by revision because the path no longer
-resolves (`git show 1cffcaaf --stat`) — is what produced `cmdkit` and
+resolves (`git show 1cffcaaf --stat`), is what produced `cmdkit` and
 router-based dispatch.
 
 **And the surface re-accreted.** Five methods on the figaro socket in May;
@@ -42,7 +42,7 @@ today:
 
 The tightening was done once and has not been done since. This plan is the
 second pass, and its job is to leave behind a *rule* that makes a third pass
-unnecessary — the May plan left a shape, not a law, so the shape drifted.
+unnecessary, the May plan left a shape, not a law, so the shape drifted.
 
 ## 2. Where the incoherence actually is
 
@@ -57,7 +57,7 @@ Three reads exist on both sockets:
 | the durable board | `aria.form` | `figaro.form` |
 
 The daemon already routes: `rpc.MethodNeedsAgent` plus `hub.go:260` serve a
-read from the store when no agent is live. **The CLI routes too** — it picks
+read from the store when no agent is live. **The CLI routes too**, it picks
 `acli.AriaPage` or `fcli.Read` at the call site. So a client must know whether
 an agent is live, which is the one fact the daemon exists to hide, and the
 routing rule lives in two places that can disagree.
@@ -74,12 +74,12 @@ Measured from the call sites today:
 
 | verb | does this shell end up attending? | where the rule lives |
 |---|---|---|
-| `fig new` / bare prompt, unbound shell | **yes** — unbind, then create-and-bind | `prompt.go:64-74` |
-| `fig form new` | **no** — "mints an unbound form" | `form_family.go:24` |
+| `fig new` / bare prompt, unbound shell | **yes**, unbind, then create-and-bind | `prompt.go:64-74` |
+| `fig form new` | **no**, "mints an unbound form" | `form_family.go:24` |
 | `fig fork` (own bound aria) | **yes**, unless `--stay` | `fork.go:185` |
-| `fig fork <someone else's>` | **no** — "a fan-out never steals this shell" | `fork.go:186` |
+| `fig fork <someone else's>` | **no**, "a fan-out never steals this shell" | `fork.go:186` |
 | `fig cast @role` that mints | **yes**, unless `--stay` | `cast_new.go:123` |
-| `fig cast` from an attended form | **yes** — mints an aria "to play it" | `cast_new.go:193` |
+| `fig cast` from an attended form | **yes**, mints an aria "to play it" | `cast_new.go:193` |
 | `fig attend <id>` | **yes**, and refuses when binding is off | `manage.go:895` |
 
 Five rules, five sites. The tell is `restoreAttendance`, whose own comment
@@ -110,22 +110,22 @@ it is Gluck's.
    *client-facing* names. The angelus keeps ONE entry point per read verb, and
    `MethodNeedsAgent` + the hub stay the only router. The CLI loses its
    `acli` vs `fcli` choice for reads.
-2. **One vocabulary.** Constants grouped by *kind* — verbs, pushes, admin —
+2. **One vocabulary.** Constants grouped by *kind*, verbs, pushes, admin,
    with the socket an implementation detail of the client, not a prefix.
 3. **Attendance as a verb outcome**, per §2.3.
 4. **A law that keeps it tight**: every new method must state which existing
    method it is NOT (the May plan's discipline, written into the plan file so
    the next accretion has to argue with it).
 
-## 4. Sequence — one idea per commit, green at every step
+## 4. Sequence: one idea per commit, green at every step
 
 | # | step | falsifier |
 |---|---|---|
-| 1 | this plan | — |
+| 1 | this plan | - |
 | 2 | attendance: strip the bind out of the mint helpers; each birth verb states its outcome | a live script that asserts the shell's binding after each of the 7 verbs |
 | 3 | collapse the read families to one client-facing surface | `fig show`/`listen`/`status` byte-identical before and after, live and dormant |
 | 4 | regroup `methods.go` by kind; delete what nothing calls | `deadcode -test` as the oracle, as W9 did (−492 lines then) |
-| 5 | slim `angelus/protocol.go` — the handler table is a table, the handlers move out | line count, and no behaviour change |
+| 5 | slim `angelus/protocol.go`, the handler table is a table, the handlers move out | line count, and no behaviour change |
 
 ## 5. Rulings (Gluck, 2026-08-21)
 
@@ -147,7 +147,7 @@ it is Gluck's.
 
 1. Does `fig form new` attend? (§2.3)
 2. Is a wire break acceptable again? The May plan said "old clients fail
-   loudly against new servers" and that is still the cheapest path — but the
+   loudly against new servers" and that is still the cheapest path, but the
    daemon and the CLI are now upgraded separately (`nix profile upgrade`
    leaves a running daemon), so a break is felt as a broken session rather
    than an error. `checkDaemonBuild` exists; does it refuse, or just warn?
@@ -157,11 +157,11 @@ it is Gluck's.
 
 | step | commit | result |
 |---|---|---|
-| 3 — one read surface | `a6388674` | the live-or-store predicate existed **three** times (per-method in `read_handlers.go`, again as `readForHub`, again in the CLI); now `routeRead` and `readFromStore`, once each. `aria.page`/`aria.context`/`aria.form` and `AriaPageRequest` deleted; `aria.read` → `figaro.ir` |
+| 3, one read surface | `a6388674` | the live-or-store predicate existed **three** times (per-method in `read_handlers.go`, again as `readForHub`, again in the CLI); now `routeRead` and `readFromStore`, once each. `aria.page`/`aria.context`/`aria.form` and `AriaPageRequest` deleted; `aria.read` → `figaro.ir` |
 | contract purity | `df03bfb0` | `api/…` depends on `api/…` and stdlib only, asserted by `api/rpc/contract_test.go`. `message`, `livedoc`, `form` moved whole; the read API's wire types split to `api/aria` |
 | sdk | `6a91fd32` | `sdk.Angelus` / `sdk.Aria`. CLI files importing a server package: **36 → 4** |
-| 2 — attendance | `d7542c16` | the mint binds nothing; `attendNew` is the one move; `restoreAttendance` deleted; `fig form new` attends |
-| 4+5 — tidiness | `3610cbd3` | `methods.go` 1,012 → 118 lines, grouped by kind, one family per file; `angelus/protocol.go` 2,011 → 371 |
+| 2, attendance | `d7542c16` | the mint binds nothing; `attendNew` is the one move; `restoreAttendance` deleted; `fig form new` attends |
+| 4+5, tidiness | `3610cbd3` | `methods.go` 1,012 → 118 lines, grouped by kind, one family per file; `angelus/protocol.go` 2,011 → 371 |
 
 ### What the instruments caught that reasoning did not
 
@@ -179,7 +179,7 @@ it is Gluck's.
 ### Open
 
 `internal/store/crashtest` failed once during the gate, under concurrent load,
-on a model divergence (`next chanLT 12, child reports 3`) — NOT the
+on a model divergence (`next chanLT 12, child reports 3`), NOT the
 `[append-lost]` lines beside it, which are documented known gaps. 20 unloaded
 runs across both branches: 0 failures. Attribution under load is in progress;
 the branch touches no store code, and the finding belongs to the store

plans/campaign-notes/README.md

@@ -5,7 +5,7 @@ Copied from `~/notes/figaro` by aria f3aa1d0b (role @980dc16c),
 finding.
 
 THESE FILES ARE CITED BY `plans/delta-seam.md` AS THE AUTHORITY FOR
-STANDING RULES — the instrument notebook and its sixteen worked
+STANDING RULES, the instrument notebook and its sixteen worked
 instances, the within-run/between-run distinction that governs which
 benchmark results are admissible, the dated REFUSED-NOT-MISSING items and
 their thresholds. Until this commit they lived only in a home directory,
@@ -15,7 +15,7 @@ WHY THEY WERE NOT VERSIONED, DIAGNOSED RATHER THAN ASSUMED: `~/notes` is
 not a git repository. Git commands run there walk UP and resolve to
 `$HOME`, which carries a bare repository's layout at its top level (HEAD,
 config with core.bare, objects/, refs/, description), created
-2026-08-16. It holds ZERO refs and ZERO objects — an empty repo, almost
+2026-08-16. It holds ZERO refs and ZERO objects, an empty repo, almost
 certainly `git init --bare` run once in the wrong directory. So `git add`
 in `~/notes` does not create a repository there; it fails with "this
 operation must be run in a work tree". That stray repo also means every
@@ -27,7 +27,7 @@ directory, and the notes directory holds personal material beside these
 files.
 
 STATUS OF THIS COPY: a SNAPSHOT, not a move. The originals remain
-authoritative and are still being edited by the arias that own them —
+authoritative and are still being edited by the arias that own them,
 each note names its author. Whoever merges this branch should decide
 whether the repo copy becomes canonical or is refreshed; a copy that
 silently diverges from its original is the same defect as a comment that

plans/campaign-notes/addtext-quadratic.md

@@ -72,22 +72,22 @@ They then checked the source and withdrew it; confirmed here independently in
         b.buf = buf
     }
 
-`String()` is **zero-copy** — it aliases the buffer. A later `Write` that grows
+`String()` is **zero-copy**, it aliases the buffer. A later `Write` that grows
 reallocates, leaving any previously returned string pointing at the old array
 with its own length and contents intact; one that does not grow appends beyond
 that string's length, which it cannot see. So materialize-on-read is cheap and
 the per-frame cost is near zero.
 
 **THE REAL HAZARD IS ALIASING, NOT COST.** A Builder-backed `Content.Text`
 handed to the composer aliases a buffer the drain loop is still appending to,
-and this codebase already has a law about precisely that shape —
+and this codebase already has a law about precisely that shape,
 `cached_log.go`: *"logView is the published window: immutable once stored, so a
 reader takes it with one atomic load and holds no lock at all."* That is what
 whoever takes this must test. The per-frame arithmetic is not the question;
 whether a held view can be mutated underneath its holder is.
 
 Still: **whoever fixes it prices the destination path first.** The warning is
-sound as a discipline even though this particular instance of it dissolved —
+sound as a discipline even though this particular instance of it dissolved,
 this campaign has been burned by a fix that looked free because nobody measured
 where the cost went. (091d162e and 6defe6f9, 2026-08-18.)
 
@@ -115,8 +115,8 @@ event is `mergeTurnTools`' map, made once per event."* **That label is wrong.**
 all.** The linear per-event term is `partialAssistant`'s `out.Content` append,
 which does escape; the tool-axis term is `toolsFromAssistant`'s append.
 
-**The measured totals are unchanged** — ~1.0 allocation removed per event,
-~38 per tool, no per-byte component — and the three-mechanism account still
+**The measured totals are unchanged**, ~1.0 allocation removed per event,
+~38 per tool, no per-byte component, and the three-mechanism account still
 covers all three axes with no residue. Only the NAME on the linear term was
 wrong.
 
@@ -127,8 +127,8 @@ concluded the MODEL was wrong when only the LABEL was.
 **THE RULE THIS COST, now standing: when a prediction is about allocations,
 ASK THE COMPILER BEFORE PREDICTING.** `-gcflags=-m` costs seconds. Reading the
 source tells you what is *written*; escape analysis tells you what is
-*allocated*, and this campaign spent four predictions — three of the
-executor's magnitudes and one of my labels — learning that those are different
+*allocated*, and this campaign spent four predictions, three of the
+executor's magnitudes and one of my labels, learning that those are different
 questions.
 
 
@@ -137,15 +137,15 @@ questions.
 **`-gcflags=-m` answers WHICH LINES allocate. It says nothing about HOW MANY
 TIMES they run.** The unverified fixed ~35 term in the stage-5 decomposition is
 a CALL-COUNT question, not an escape question, and reaching for escape analysis
-to settle it would be the right tool on the wrong shape — the same class of
+to settle it would be the right tool on the wrong shape, the same class of
 error as reading `make(map...)` and calling it an allocation because it is
 spelled like one.
 
 What would settle it: `-memprofile` with `alloc_objects` attributed by call
 site, or a differential count with one call site removed.
 
 6defe6f9's hypothesis, recorded as a hypothesis: before stage 5, `noteAssistant`
-had TWO call sites —
+had TWO call sites,
 
     turn.go:519  a.noteAssistant(asmMsg.message())   per BUS EVENT  -> the linear term
     turn.go:523  a.noteAssistant(&staged.Payload)    per ROUND      -> fixed-term shaped