fix: handle empty SMW ask results in semantic_search

[LukasGold]

Fixes the crash behind #145 and explains the "random results" in #111.

Root causes

Three separate defects, all reproduced live against healthbatt.projects01.open-semantic-lab.org.

1. Zero-result queries raise AttributeError (fixed here)

SMW's action=ask serialises a non-empty result set as a JSON object keyed by page title, but an empty one as a JSON array. semantic_search called .values() on it unconditionally (src/osw/wiki_tools.py:257), so every zero-result query raised AttributeError: 'list' object has no attribute 'values' instead of returning [].

[[HasUuid::b5cbec50-fb7c-4c65-b208-2a20c6c32fb8]]  -> results: []   (list)
[[HasLabel::BMD_StandardCU_HealthBatt]]            -> results: {..} (dict)

Affects single queries, sequential batches and parallel batches. return_json=True happened to survive because it returns before that line. OSW.query_instances() (src/osw/core.py:2007) is affected too, so querying a category with no instances raised instead of returning an empty list.

2. Silent truncation at limit (surfaced here)

SearchParam.limit defaults to 1000 and there is no pagination, so a larger match set is silently reduced to an arbitrary subset. Measured on healthbatt: [[Category:Item]] matches 2637 pages, semantic_search returned 1000 with no indication.

3. exists flickers server-side (surfaced here, not fixable in this library)

The exists == "1" filter silently discards results. The field only ever takes '1' or '', but it is not stable: across 8 identical calls, 8 pages flipped between the two, and no page was consistently ''. The flips are perfectly correlated across pages (calls 0,2,3,4,7 agree; calls 1,5,6 agree), which points at inconsistent state between backends or an SMW query-result cache rather than at real page deletions. Consecutive identical calls returned 998 or 994 titles.

This is the "random results" from #111. It needs to be addressed on the OSL instance side; this PR only makes the loss visible.

Changes

• _ask_results_as_dict() normalises the ask payload to a mapping, so an empty result set yields [] instead of raising
• warnings.warn when the result count meets the requested limit, emitted before the return_json early return so both return modes get it
• warnings.warn when entries are dropped by the exists != "1" filter, naming how many
• 7 unit tests covering zero results in both return modes, sequential and parallel batches with an empty query, both warnings, and the no-warning case

Filter semantics, the default limit, and prefix_search are unchanged.

Verification

• pytest tests/ --ignore=tests/integration -> 61 passed, 1 skipped
• New tests confirmed failing first with the expected AttributeError / DID NOT WARN
• Re-ran the live probes against healthbatt: sequential batch, parallel batch and return_json with a zero-result query all return results now instead of raising; both warnings fire with real counts

Out of scope, noticed while probing

single_query += f"|limit={query.limit}" is appended unconditionally, so a caller-supplied limit in the query string is silently overridden: [[Category:Item]]|limit=2 is sent as [[Category:Item]]|limit=2|limit=1000 and SMW honours the last one. Verified live (returned 1000, not 2). Not changed here.

[github-actions]

Release preview

Merging this PR would release v2.0.3 (current: v2.0.2).

Changelog preview (truncated)

## v2.0.3 (2026-08-31)

### Bug Fixes

- Handle empty SMW ask results in semantic_search
  ([`e00ca8b`](https://github.com/OpenSemanticLab/osw-python/commit/e00ca8be80eb1575a520b609d5d988604ad59e24))

### Testing

- Rename oold.py to oold_test.py so its tests are collected
  ([`20072a9`](https://github.com/OpenSemanticLab/osw-python/commit/20072a9249cd97126a222c62a70f84e0433343ef))

_{Preview via python-semantic-release and conventional commits.}