[ddehilster]

NLP++ runs as a C++ extension, so a grammar whose loop never ends can't be stopped from outside. A Python caller has no thread to interrupt, only a process to kill. And a loop that makes a new string every turn keeps growing until the process is killed for its memory. One such loop took a document extractor down five times in half an hour on one document, along with every analysis queued behind it. The loop itself came from the string-position bug fixed in #799; this PR covers the next loop, whatever causes it.

API

• NLP_ENGINE::setRunLimit(ms): a limit on each text's analysis. 0 (the default) means no limit.
• NLP_ENGINE::lastRunStopped(): true after an analyze() / analyzeStream() call that stopped a run at the limit.
• nlp -TIMELIMIT <seconds>: a run it stops exits 2. That's distinct from success (0) and from the missing analyzer or input that exits 1.

How it stops

• Each Parse takes the limit when it's created. Parse::overBudget() is checked on every while turn (it reads the clock every 256th time) and before every pass.
• Past the limit, the loop fails its region and no further passes run. The stop is reported once, in err.log and on stderr, naming the pass and its file:
  [Run stopped: past its time limit of 2000 ms in pass 3 (loop); the rest of the analysis was not run.]
• The run's string table is freed afterwards as usual, so the memory comes back.
• Compiled analyzers get the same check: the generator writes Arun::overbudget(nlppp) into every while condition.

Tests

• New fixture run-limit, run on Windows, Linux and macOS. It loops forever in pass 3, and checks that pass 2 ran, that neither the line after the loop nor pass 4 ran, that the message is on stderr, and that the exit status is 2.
• The production analyzer that hung, run on 4.2.39's string functions so it still loops: -TIMELIMIT 5 stops it at 5 s with exit 2.

No limit means no change

• 881 fixture inputs of 17 production analyzers give byte-identical output.json to 4.2.39.
• 3 million while turns take the same time (about 3.0–3.2 s) with no limit and with a 600 s limit.
• 52 of the 54 Windows CI steps pass locally. The two that don't need CI's environment: compile libraries for compiled-bridge, and the analyzers checkout for test-nlp.

This is rebased on #799 and builds as 4.2.41. The str-positions and run-limit steps both pass on the combined build.

The NLPPlus side (analyze(..., timeout=)) follows once this is in a release.

🤖 Generated with Claude Code