feat(server): session-owned pinned geometry (run --size, pty resize)

[schickling-assistant]

Restores the session-owned geometry half of #92, which #96 narrowed out when it landed only the client-role half. Problem statement and evidence: #95.

#96 shipped attach --no-resize, ATTACH_FLAG_GEOMETRY_NEUTRAL, capabilities.geometryNeutralAttach and clients.geometryNeutral. Main still has no run --size, no pinned session geometry, no pty resize, no RESIZE_FLAG_AUTHORITATIVE, and no geometry block in stats. This PR adds those, ported onto current main rather than cherry-picked (main moved 34+ commits and the surrounding code changed).

Why this is not redundant with #96

1. Neutral attach does not protect a running TUI. --no-resize lets a client opt out of size negotiation. It does nothing about a plain, non-neutral attacher, which still drags the session to its own dimensions via min-wins negotiation. Only a pinned session revokes the vote from every client. The two are complementary: neutral attach is opt-in on the observer, pinned geometry is enforced by the session.

2. --size is load-bearing for an existing downstream consumer. A consumer invokes pty run -d -a --size 160x48 --id <name>. On main --size is unknown, and the run flag loop (src/cli.ts) ends in else break — so parsing halts at --size and --id is never parsed. The session silently gets a random id instead of its intended name, and session identity is load-bearing there. This is a silent failure, not an error. There is now a regression test for exactly this ordering.

3. fractal (a TUI agent-IDE, shipped) is fully degraded without it. It gates its non-disturbing attach on a top-level geometry block in stats. Main never emits one, so every session probes as incapable and fractal silently falls back to read-only attach — total feature loss even though main's neutral attach works.

4. A fixed render size prevents the 80x24 TUI-freeze class of bug for headless / no-tty agent sessions.

What's in this PR

• pty run --size <cols>x<rows> — pins the session's geometry at spawn.
• pty resize <ref> <cols>x<rows> — authoritative resize on a running session; pins it and emits exactly one SIGWINCH (no redraw nudge, since this resize is legitimate).
• RESIZE_FLAG_AUTHORITATIVE / encodeResizeAuthoritative in protocol.ts.
• negotiateSize() is inert while a session is pinned, so plain attachers cannot reflow it.
• Stats gains geometry: { owner, pinned } and a capabilities.pinnedGeometry flag.
• src/completions.ts (SSOT) covers resize and run --size; committed completions/pty.{bash,fish,zsh} regenerated from it.
• New tests/session-geometry.test.ts (12 tests).

Compatibility with what #96 landed

Main's surface is added to, never reverted. capabilities.geometryNeutralAttach, clients.geometryNeutral, decodeAttachFlags and the --no-resize spelling are unchanged, and no --neutral alias is introduced (no consumer uses it).

RESIZE_FLAG_AUTHORITATIVE reuses the same optional byte-4 flag slot as ATTACH, so a legacy 4-byte RESIZE frame stays byte-identical and old daemons keep parsing it. There is a test pinning that.

Note on capability advertisement

Support is advertised both ways, on purpose:

• capabilities.pinnedGeometry: true is the clean contract and the recommended thing to probe.
• The geometry block is emitted unconditionally, because fractal currently sniffs block presence. Keeping it means fractal works against this build with no change; the capability flag is the migration path.

Verification

Run locally in a clean worktree:

• npm run typecheck — clean.
• npm test — 1397 passed, 2 failed. Both failures (shells > zsh, screenshot > vim) reproduce identically on unmodified 0a9be8c and are environmental (no zsh on this machine); confirmed against a clean checkout of that commit.
• npm run verify-docs — 12 passed, 1 failed (doc example 12, a vim example); also reproduces identically on unmodified 0a9be8c.
• nix build .#pty — succeeds.
• Manually exercised against the nix-built binary: run --size (confirmed --id after --size is honored), a plain 80x24 attach against a pinned 160x48 session (session stayed 160x48, client reported as non-neutral), pty resize, pty stats --json, pty completions fish.

New tests cover: run --size parses and pins (including the --id-after---size regression); a plain attach does not reflow a pinned session; a plain attach does still reflow an unpinned one (default behavior untouched); pty resize emits exactly one WINCH; stats exposes geometry for both pinned and unpinned sessions; and legacy-frame byte compatibility.

One note on the WINCH test: the baseline is not zero, because attaching a client triggers the existing redraw nudge (resize -1 col and back), which is itself two real WINCHes. The test measures the delta the authoritative resize adds, which is 1.

Deviation from scope, flagged for review

tests/help.test.ts's no-drift check was already red on main: the completions dispatch case merged in #106 was never added to its coverage lists. Since this PR adds a new subcommand to that same list, I fixed the pre-existing failure in passing — which required capitalizing usage: to Usage: in the completions help synopsis to match the repo-wide Usage: pty ... contract, updating the assertion in tests/completions.test.ts, and adding a pty completions line to the top-level help. Happy to split that into its own PR if you would rather keep this one narrow.

No nix/packaging changes — this is a source-only PR.

[schickling-assistant]

Closing unmerged: the consumer need this PR was written to serve turned out not to exist.

• The downstream TUI (fractal) gated its non-disturbing attach on a top-level geometry
  block in pty stats. Canonical pty emits no such block and no clients.neutral, so that
  code path was provably unreachable — enter already ran a plain attach at pane size.
  The dead branch has been removed downstream, with no change in attach behaviour.
• The 80x24 symptom originally cited as motivation was a defect in the consumer (it launched
  the attach child over pipes, so the child reported the fallback geometry). Fixed there by
  giving the child a real PTY. It was never a pty gap.
• attach: stop nudging a child that is already the right size; remove --no-resize #109's nudge-skip (nudge only when geometry actually differs) addresses the one real defect
  in this area, generally and with a much smaller surface than pinned geometry.

The only remaining consumer of run --size was a headless launch wrapper, which has been
updated to drop it. That wrapper now starts sessions at the daemon default rather than a
pinned size — an accepted, measured tradeoff, not an oversight.

Reopening this would need concrete evidence that headless sessions rendering at the daemon
default causes real trouble, rather than the speculative case made here.