fix: refuse an over-long state before tokenizing every prompt; sturdier tests and release

[feder-cr]

Fixes from a review of the code, the tests and the docs (three parallel reviews, each finding checked against the code before fixing).

Server

• Over-long state: a state longer than the context made every prompt too long, but all prompts were tokenized first, each holding the whole state. A 250 KB state with 1,024 questions (or 40 choices of 26 options) took every core and gigabytes of memory, and an allocation failure on a tokenizing thread ended the process. The state is now measured first and the request refused with the same error as before (the first question's prompt length): 422 in 0.22 s, memory unchanged (1.03 GB peak).
• Token buffers: tokens no longer keep a buffer sized to their text.
• Thread errors: an error on a tokenizing thread fails its request instead of std::terminate; an error during idle warm-up stops the warm-up instead of the server.
• Score probabilities: a score level's probability is never below 0 (pooled means could round to −1e-17).
• Duplicate snapshots: two batched requests with the same new state keep one snapshot, not two.

Tests, CI, release

• check.py:
  • refuses to run if its port already answers (it used to test that server instead of the build);
  • stops a server that did not start, and waits by time rather than by tries;
  • puts a timeout on every request and resolves paths from where it was started;
  • does not pass JEV_API_KEY on;
  • makes the choice and score checks independent of step order.
• score-test: 20,000 random threshold sequences, no level below 0, levels sum to 1.
• Workflows: jobs have timeouts. release.yml removes SHA256SUMS.txt before replacing the archives and attaches it again once all three are in, so a failed or partial run never leaves a stale one.

Docs: about 40 README and wiki pages were stale:

• they still said score is refused;
• they described the 8,192-token context as per request (it is per question's prompt);
• they said every response carries Server-Timing (only successful ones do);
• the offline page left out the build's pip install and its network step.

The options table gains --name, --ctx and --warmup.

Verified locally: full check.py --exact --threads 8 green, with zero-tolerance parity 215/215 identical (no numerical change).

🤖 Generated with Claude Code