docs: the roadmap and AGENTS.md catch up with #102, #328 and #329

.roadmap/features/f-builtin-terminal.md

@@ -66,6 +66,9 @@ the mouse back for a local selection. The copy chord used to fall back to a
 cache of the last text copied when no selection was visible, which in a pane
 the child owns — never a visible selection — overwrote the child's own copy on
 every chord (#316); the cache is gone and the chord asks the terminal whether
-it holds a selection, scrolled out of view or not. One contract bug remains on
-the surface: the `emitted_lines_never_drift` property has a known failing
-scroll sequence whose seed was never committed (#102)
+it holds a selection, scrolled out of view or not. The scroll-drift property no
+longer carries a known failure. This entry used to call its failing sequence
+the one contract bug remaining on the surface (#102). It was f32 noise in the
+test's own reference sum, not a scrolling defect, and `b1d8045` had already
+fixed it the day after it was filed. The issue closed on 2026-09-24, and its
+seed now replays (#328)

.wrap.md

@@ -362,3 +362,14 @@ upstream" and moves to the next one. A conflict there is resolved as usual, but
 the `--continue` again needs the user's terminal for the signature — plan for
 two hand-offs per rebase, not one. Seen rebasing #308 after six merges had
 landed under it.
+
+**`markdownlint-cli2`'s summary counts files with findings, not files linted.**
+A clean run ends `Summary: 0 issues in 0 files`, which reads as "linted
+nothing". Coverage is the line above it, `Linting: N files`. The repo config
+adds its own `**/*.md` globs to whatever path you pass, so N is ~100 from the
+root or from any worktree. In the #102 pass that line got cut off by a
+`tail -1`, the summary was read as a vacuous pass, and a wrong "markdownlint
+lints nothing from a worktree" claim went to the user and nearly into this
+file. Read the
+`Linting:` line, or inject a known violation once. A single long token will not
+do, because MD013 skips what cannot wrap: use a long line with spaces.

AGENTS.md

@@ -298,36 +298,24 @@ carry is modifiers: the event lost its `modifiers` field in #311 as dead
 weight, and it comes back as one type shared with `KeyMods` when a caller
 needs it, not before.
 
-**Still open.** Three features and two defects: `F-mcp-agent-loop` (#196 —
-below), `F-mcp-attach` (#267 — the bridge is reachable only from a session
-termherd spawned, so the launcher itself cannot drive it), the pointer at
-termherd's own chrome (#301 — the sidebar, tabs and gutters still have no
-mouse), `enter` on the two renames (#246), and a doc editor that discards
-unsaved edits when it closes (#248). None of the five blocks another. A third
-defect, the copy chord overwriting a mouse-mode program's own clipboard write
-with termherd's last copy (#316), is fixed: the shell keeps no copy cache any
-more, and asks the terminal whether it holds a selection at all.
-
-`F-mcp-agent-loop` (#196 — the composed prompt→wait→read in one
-round trip) is a child of the #90 epic — no longer the last one, since three
-siblings joined it. With `screenshot`, the keyboard tools and the terminal
-pointer reaching the child, the capability reads as whole in three parts:
-drive the UI, see the pixels, read the terminal. #196 *composes* the wait,
-which #236 had to fix first — building it on a synchronisation that never
-fired would have been building on sand, and that ordering constraint is now
-discharged.
 **Still open.** Two features and two defects: `F-mcp-attach` (#267 — the bridge
 is reachable only from a session termherd spawned, so the launcher itself cannot
 drive it), the pointer at termherd's own chrome (#301 — the sidebar, tabs and
 gutters still have no mouse), `enter` on the two renames (#246), and a doc editor
 that discards unsaved edits when it closes (#248). None of the four blocks another.
+The list leaves out the bridge's hardening defects (#284, #293). A third
+defect, the copy chord overwriting a mouse-mode program's own clipboard write
+with termherd's last copy (#316), is fixed: the shell keeps no copy cache any
+more, and asks the terminal whether it holds a selection at all.
 
 **`F-mcp-agent-loop` shipped in #196 (`prompt_in_session`).** It composes
 `run_in_session`, `wait_for_status`, and `read_terminal` into a single MCP round
 trip, with bounded step-by-step execution and opt-in protection for nested Claude
 sessions. `F-mcp-agent-loop` was a child of the #90 epic — with `screenshot`,
 the keyboard tools, terminal pointer, and composed agent loop shipped, the live
-bridge control surface is complete.
+bridge control surface is complete. It could only be built after #236: a
+composed wait on a synchronisation that never fired would have settled by
+timeout every time.
 
 **Every overlay can now be left from the keyboard** (#237). An open sidebar
 session-rename used to swallow every key including `escape`, parking the whole
@@ -558,6 +546,17 @@ exists). Do not relax them locally.
   mid-run (see `.wrap.md`): success and vacuity are indistinguishable from the
   outside. Probe with the operation you actually intend to perform, and if that
   is destructive, assert on something the operation *must* have changed.
+- **A proptest seed file is keyed by the test's source path, so moving the
+  test orphans it without a word.** proptest's default persistence reads
+  `proptest-regressions/<path of the .rs file>.txt`, derived from `file!()`.
+  When the scroll-drift property left `shell/terminal.rs` (for
+  `shell/terminal/mod.rs`, then `shell/terminal/canvas.rs`), its two committed
+  seeds stopped replaying from 2026-07-12 to 2026-09-25, and nothing reported
+  it: an unread seed file gives no signal either way. When a refactor moves a
+  file holding a `proptest!`, `git mv` its regressions file in the same commit.
+  The file is test input to CI as well: the `rust` path filter now watches
+  `proptest-regressions/**`, since the PR that moved it came up green with
+  `test` and `portable` both skipped.
 - **An unfinished check is not a passing check, and reporting it once is not
   reading it.** `portable` is the only gate that runs a Windows *behaviour*
   before merge, and it takes minutes while every other job takes seconds — so

ROADMAP.md

@@ -259,9 +259,12 @@ the mouse back for a local selection. The copy chord used to fall back to a
 cache of the last text copied when no selection was visible, which in a pane
 the child owns — never a visible selection — overwrote the child's own copy on
 every chord (#316); the cache is gone and the chord asks the terminal whether
-it holds a selection, scrolled out of view or not. One contract bug remains on
-the surface: the `emitted_lines_never_drift` property has a known failing
-scroll sequence whose seed was never committed (#102)
+it holds a selection, scrolled out of view or not. The scroll-drift property no
+longer carries a known failure. This entry used to call its failing sequence
+the one contract bug remaining on the surface (#102). It was f32 noise in the
+test's own reference sum, not a scrolling defect, and `b1d8045` had already
+fixed it the day after it was filed. The issue closed on 2026-09-24, and its
+seed now replays (#328)
 
 <a id="f-close-confirm-policy"></a>