llauncher v0.5.1-alpha

Tag v0.5.1-alpha on ca8314e. Ledger #526; measurement record docs/experiments/2026-08-27-ui-stall-quantum.md.

[0.5.1-alpha] — 2026-08-27

Windows UI performance — the change 0.5.0-alpha deferred, delivered

0.5.0-alpha shipped the PID-first primitives (#466 Phase 1) with the note "no user-visible latency change in this alpha; that's Phase 2." This release delivers the change, by a different route: the stall was a Windows envelope cost, not the architecture.

• Root cause (#521). Every psutil process walk fetched cmdline() for all ~300 processes — ~3.1 s per walk on Windows (a per-process handle open + PEB read with no batch path), milliseconds on Linux. Every Streamlit stall on the Windows seat was an integer multiple of that quantum; #497/#498 had already removed the walks from the title-paint and tab-switch paths (≈40 s → ~0 each), but state.refresh_running_servers and the uncached is_port_in_use (via can_start, every Start click) still walked.
• Fix (#522). Walks read pid+name and call cmdline() only for name-matched candidates (configured llama-server binary or an interpreter/shell, so wrapper launches still match). is_port_in_use reads the TCP socket table — LISTEN only, bind-probe fallback on AccessDenied — instead of scanning every argv for --port. find_available_port reads the socket table once. Envelope only: the OS remains ground truth (ADR-LLNCH-008 rule 3); the TTL cache and state.py are untouched.
• Measured (Windows seat, headless driver, same protocol before/after). Models tab clickable 50.0 s → 1.7 s; port edit → Start enabled 43.4 s → 0.59 s; UI-reports-running lag +32 s → eliminated (the row inverted by 1.5 s — a driver detection nuance, not a real lead); tab switch / expander 33–46 ms. Microbenchmark (same run, docs/experiments/bench_psutil_walk.py): full walk 3,162 ms → name-first 2.4 ms; is_port_in_use ~6,170 ms → 0.8 ms. Operator-confirmed indistinguishable from the Linux UX (recorded on #521). Record: docs/experiments/2026-08-27-ui-stall-quantum.md (+ bench_psutil_walk.py).
• Behavior change. "Port in use" now means a socket holds the port, not "some process's argv mentions it" — strictly more correct, and TIME_WAIT no longer reports a just-stopped port as occupied.

Also since 0.5.0-alpha

• Handoffs: four Windows-seat docs (2026-08-26, 2026-08-26b, and its two addenda)
• fix(remote): per-verb client timeout for start/swap so slow model loads are not reported as failures (#503) (#511)
• perf(ui): remaining st.rerun() sites cost one script run per click (#498) (#510)
• docs(readme): correct run.sh/run.bat paths, drop stale agent-bg (#501) (#508)
• test(mcp): isolate lazy-singleton tests from real psutil scans and nodes.json (#496) (#506)
• fix(process): name the config entry in the extra_args deny-list error (#462) (#505)
• perf(ui): hoist state.refresh() to one call per script run (#497) (#507)
• fix(ui): Edit / Save Changes run the script once — on_click callbacks replace st.rerun() (#494) (#499)
• docs(adr): namespace bare ADR-026 citations in source comments (#489) (#492)

Known issues (carried or new)

• 57 environment/portability test failures on the Windows seat at d433c75 (#523, supersedes #364); CI-green still requires Linux.
• stop reports success before the port is released (#518) — reproduced live 2026-08-27; #515–#517 (stop/start mechanics) remain open.
• #466 Phase 2/3 (PID-first state.py) remains open; no longer load-bearing for Windows latency.
• llama-server discovery is case-sensitive/substring on name and argv (#524, fix in flight).
• Restart-Service llauncher-agent terminates managed llama-server children (#480, carried from 0.5.0-alpha).