feat(cli): claw run hands the run to the server, starting one if needed

[EdmondDantes]

What

The second slice of TODO item 2. claw run stops opening a project db and running the solver in its own process; it becomes a thin client that hands the run to the workspace's server — the one process allowed to write.

• New ServerClient (src/ServerClient.php):
  • running() — a live server's address, or null.
  • ensure() — that address, starting a server when none answers and waiting for it to come up.
  • startRun() — POST /api/projects/{key}/issues/{id}/start, mapping 202 → ok, 409 → "already active", anything else → an error the CLI surfaces.
• runIssue() now resolves the issue, then ensure() + startRun() — no Config, no agent, no IssueRunner in the CLI process.

Autospawn (cross-platform)

When no server answers, the CLI starts one. The launch command is rebuilt from PHP_BINARY and CLAW_SERVER_EXTENSION (the server extension, defaulting to true_async_server.so; on Windows a .dll, so set the env var), because the CLI itself runs without the server extension.

spawnCommand() returns an argv list, not a shell string, so it carries no quoting to platform-mismatch. launchDetached() then branches:

• POSIX — setsid --fork gives the server its own session and returns at once (without --fork, setsid execs the server in place and proc_close blocks for its whole life — found by the CLI hanging in the live test).
• Windows — proc_open with create_new_console + bypass_shell, cutting the server from the CLI's console so it survives the exit.

Either way stdio is pointed at files (/dev/null or NUL) so the server never inherits — or blocks — the pipe the CLI was invoked through. The loser of a bind race exits on the duplicate-guard from #91, and the readiness poll finds the winner.

Verified live (POSIX)

Against the real server extension on a Linux/WSL machine:

• cold start — no server → starting one… → run launched, CLI returns in the same second;
• warm reuse — a second run finds the running server, no respawn;
• 409 — a start for an in-flight issue prints a run for this issue is already active;
• create → run — claw -i opens an issue, claw run starts it on the server (run row created, status running).

Windows is written to the API but unverified — no machine to run it on here.

Behaviour changes (intended)

• No live console trace and no terminal solver-approval prompt on claw run. The approval prompt was a human-approval gate the project does not want; the live trace returns in the next slice (poll GET …/runs/{id}/trace?since= + answer the gate over POST …/answer).
• claw run now needs a server — one running, or the server extension available to start one. This is the "server becomes required" cost item 2 records and accepts.

Not in this PR

• /start returning the run id, and the polling tail that renders the trace and drives the human gate (next slice).
• claw -i / -c still write locally; folding the remaining writers behind the server is later work.

Stacking

Based on feat/server-discovery-file (#91), which adds the ServerLocator this builds on. Retarget to main once #91 merges.

Checks

composer qa green — cs, PHPStan level 8, Testo (452 tests). New tests/ServerClientTest.php covers the start status mapping, the target URL, running() on an empty workspace, and the spawn-argv shape.