Separate a failed run from a wait that could not finish

[cfsmp3]

Follow-up to #5, whose fix was right and whose reasoning applies once more one layer over.

The same collision, one layer up

#5 established that run wait must not share an exit code with something that means a different thing. It still does. An ApiError with no code of its own exits 1, and 1 is what this command uses for a run you waited for failed:

server error (500)     -> exit 1
a run genuinely failed -> exit 1
not found (404)        -> exit 4     # mapped, so already distinct

Measured on master at fe8384a.

These want opposite responses. The gating script from #5's own description:

sp run wait $ID
case $? in 0) merge ;; 1) block ;; 9) retry ;; esac

A 500 from the platform takes the block branch — the branch is reported bad although nothing was learned about it — and retry, the correct response to a 500, is the one path the script will not take. Same shape as the original bug: a condition that is not an outcome wearing an outcome's code.

The fix

Codeless API failures exit 10. Errors that map to a code of their own keep it, so nothing that already worked changes:

condition	before	after
every run passed	0	0
a run failed or was canceled	1	1
timeout	9	9
API error mapping to a code (404, 401, 429, 409 …)	4, 6, 7, 8	unchanged
any other API failure (500, malformed response)	1	10

After this, 1 under run wait means exactly one thing: a run reported a bad verdict. Everything else that can go wrong has its own code, which is the property #5 was after.

case $? in
  0)     merge ;;
  1)     block ;;      # the branch is bad
  9|10)  retry ;;      # the platform is; the branch is unjudged
esac

Testing

Two tests, both of which fail without the change:

• test_a_broken_wait_is_distinguishable_from_a_failed_run — a 500 and a failed run must not share a code
• test_api_errors_that_map_to_a_code_keep_it — 404→4, 401→6, 429→7, 409→8 are untouched, so the remap is narrow

220 pass under both pytest and unittest discover; isort, pycodestyle, pydocstyle, mypy clean. README and AGENTS.md tables updated, including the worked case above.

Also

Drops a stray blank line that landed above constants.py's module docstring in #5. Harmless — __doc__ still resolves — but it is not meant to be there and no linter flags it.