fix: recoverable client on connection failure

[iemejia]

Split from #221 to ease review.

Replace OnceLock<Client> with thread_local RefCell<Option<Client>> that auto-resets on connection errors (broken pipe, pool timeout, connection refused, etc.). Subsequent calls transparently re-create the client pool instead of failing permanently for the session's lifetime.

Add is_connection_error() heuristic with unit tests and pg_test integration coverage.

Files: src/client.rs, src/lib.rs
Tests: 6 unit tests + 1 pg_test

[pinodeca]

Reviewed with Opus 4.8

Suggestions (non-blocking)

1. Unnecessary String allocations. The new .to_string() conversions (inst_id, fn_name, inp, evt_name, etc.) aren't needed. rt.block_on(...) runs synchronously within the function scope, so the closure can capture the original &str arguments directly. This adds an allocation per call on the hot path of df.start()/df.signal()/df.cancel(). Minor, but easy to drop.

2. is_connection_error is broad. Matching bare substrings "connection" and "closed" can produce false positives (e.g. a business error mentioning "connection" or "closed"). The consequence is benign — just an extra pool re-creation — but worth a one-line comment noting that false positives are acceptable-by-design (cheap recovery, never incorrect). Anchoring on phrases like "connection refused"/"connection reset"/"connection closed" would tighten it if you'd prefer.

3. Reentrancy caveat. The immutable borrow is held across f's entire block_on. Today none of the three callers re-enter with_duroxide_client, so it's safe, but a future caller that did would panic. A short comment on with_duroxide_client warning against reentrant calls would protect maintainers.

4. Removed set_var unsafety note. The original comment warning that std::env::set_var becomes unsafe in the 2024 edition was dropped. The crate is edition = "2021" so this is fine today, but that note was a useful landmine marker for a future edition bump — consider keeping it.

5. Test redundancy. The pg_test (test_is_connection_error_detects_failures) largely duplicates the #[cfg(test)] unit tests. Not a problem — just flagging that the is_connection_error_for_test wrapper exists solely to expose the fn through the pg_test harness.

Verdict

Approve with minor nits. None of the above is blocking; #1 (drop the allocations) and a comment for #2/#3 would be the highest-value follow-ups.