Detach async jobs by forking instead of SIGKILLing the web server (NEA branch)

[bjfultn]

Port of #28's two code fixes onto the NEA branch, so NEA has a branch to test
against. Branched from hotfix/table-validation at bbc1108 (the #26 merge),
so it already carries the VOSI /capabilities and /availability work.

TAP/tap.py only — no tests, matching this branch's convention. The test
coverage for these fixes lives on #28 against develop.

What changed

1. Async submit forks instead of killing the web server.
__printAsyncResponse__ used to send the 303, sleep two seconds, then
os.kill(os.getppid(), SIGKILL). Under CGI that parent is the web server child
serving the request. Killing it truncates the response, and behind a reverse
proxy it leaves the proxy holding a dead upstream connection — which the proxy
reports as 5xx on the next request or two routed over it.

The submit path now forks: the parent publishes the job, sends a framed 303 and
exits normally; the child setsid()s out of the request's process group, points
its stdio at /dev/null, waits on a pipe until the parent has published the
status document, then runs the query. uws:runId carries the child's pid, since
ABORT signals whatever runId names.

2. ABORT and unrecognized phases no longer run the job.
Every phase fell through one shared response path and then into the query, so
an aborted job reported ABORTED and returned results anyway. Terminal phases
now write their status, answer the client, and sys.exit(). Only RUN detaches.

Verification

Ran the full async flow out of process against this branch, with a SQLite
fixture (no server, no proxy):

1. harness process survives the run        alive
2. submit async job                        303, jobid issued
3. POST PHASE=RUN -> poll                  303, phase COMPLETED
4. GET results                             200 (494B)
5. POST PHASE=ABORT                        phase ABORTED, result files: 0
6. VOSI regression check                   /capabilities 200, /availability 200
                                           (one HTTP response each)

Same harness against this branch's base (origin/hotfix/table-validation) dies
at step 3 with exit 137 — SIGKILL — which is the bug: the CGI killing the
process that invoked it. That is the red half of the check; the fix is the green.

Step 5 guards the second fix: before it, an aborted job still produced a result
file.

Notes

• The four other hand-rolled 303 See Other emitters in tap.py are untouched
  here, exactly as in Detach async jobs by forking instead of SIGKILLing the web server #28. Only __printAsyncResponse__ gained Content-Length
  and Connection: close.
• This branch does not carry Fix malformed HTTP responses from /capabilities and /availability #27's test suite or ruff/CI plumbing, which
  don't exist on hotfix/table-validation.

🤖 Generated with Claude Code