PR #25453 - manaflow-ai/cmux

Summary

Part 1 of 3 of #24281, split to stay under about 800 changed lines. The code and tests are #24281's at da7658bf (which had an independent Codex SHIP there), with only the cut points chosen.

About 25 Codex reviewers started with a raw acpmux new over reach went around lane-place's cap of 12, filled huge-blue's disk and pushed its load to 50-80. Every new session goes through the daemon's create, so the daemon now refuses one when:

~/.cache/worktree-reclaim/under-floor exists (the disk is under its free-space floor), or
this host's live sessions are at lanes.max_lanes for this machine in the fleet registry (FLEET_REGISTRY, else ~/.ssh/config.d/fleet-registry.json, else leo-workspace's machines/fleet-registry.json).

What is checked:

In this part: session/new (no adopt, or an adopt that may fork), session/fork and the composer's aside (_acpmux/ask).
Never checked: sends, resumes, restarts, side chats, handoff targets and adopts that never fork.

How it behaves:

An admitted create counts as live until it ends, so concurrent creates at the cap cannot all pass.
A create refused while another create holds its name says the name is taken, so new --reuse takes the winner.
_meta.acpmux.forceAdmit, from this host's socket or app only, admits past a refusal. So does the person's own new chat. Both are logged.
With no registry, or no lanes entry for this machine, every create is admitted, so acpmux outside the fleet is unchanged.

What each later part adds:

Part 2: the CLI side (--force-admit on new, fork and new --parent, exit 5 on a refusal), sub-session starts, and the CLI integration tests (tests/admission.rs, tests/admission_subsessions.rs).
Part 3: adopts. An adopt of a conversation a session holds attaches and needs no room. One whose holder ends first asks for room. Concurrent adopts of one conversation converge (admit_adopt, NewRequest::attach_only, tests/admission_adopt.rs).

Testing

Run on huge-blue with a scratch HOME, because this host's own under-floor flag is set and the daemon would otherwise read it:

cargo test -p acpmux --lib: 981 passed, 30 failed. Base 2815ee94 fails the same 30 here (temporary-folder, store-permission and TUI tests that depend on this host's TMPDIR) and passes 972. The 9 extra passes are the new hub::admission tests.
cargo test -p acpmux --test hub_integration: the same failures as base, plus stalled_turn::a_turn_counts_from_its_prompt_being_written once. It then passed 4 of 4 reruns.
subsessions, agent_tools, local_app_trust_gate, remote_guard and session_env pass.
send_attach_family and remote_guard_modes fail the same way on base.
cargo fmt --check and cargo clippy -p acpmux --all-targets -- -D warnings pass.

Observed / Not run

Observed: the runs above.
Not run here: the CLI integration tests of the refusal itself. They land with part 2, because they drive --force-admit and exit 5.

Changelog

Added: acpmux refuses a new agent session while its host's disk is under its free-space floor or the host is at its fleet lane cap; --force-admit overrides