Add conversation_id and response_id fields to ChatResult for tracing
eval conversations back to the server. Failed conversations can be
preserved for post-mortem inspection via --preserve-failed.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

_args_match checked emit_widget (renamed to user_requested_search in the
tool schema) and limit (optional, server-side default). Both mismatched on
every real model call, so tool_correctness was always False regardless of
whether the model used the right keywords and object types.

Fix: evaluate only keywords (case-insensitive) and object_types — the two
fields that actually determine whether the search was semantically correct.

Bump development and test tooling to current releases:
- ty 0.0.40 -> 0.0.55
- ruff 0.15.15 -> 0.15.20
- pre-commit 4.0.1 -> 4.6.0
- tox 4.32.0 -> 4.56.1, tox-uv 1.29.0 -> 1.35.2
- vcrpy 8.0.0 -> 8.2.1 (root + sdk/pandas/fdw test groups)
- moto 5.1.22 -> 5.2.2 (re-lock within existing >=5.1.6 spec)

ty 0.0.55's improved inference flags the cast in MetricValueFilter.description
as redundant, so drop it (and the now-unused typing.cast import). All other
runtime deps left capped as-is; urllib3 stays pinned to the OpenAPI generator.

Validated: make lint, format, type-check green; sdk/pandas/fdw/pipelines tests pass.

jira: trivial
risk: low

Apply ruff autofixes surfaced by `make lint-fix` across gooddata-eval:
remove unused imports and re-sort import blocks, and add the missing
`# noqa: PLC0415` to the deliberately function-local `timezone as _tz`
imports (matching the `_dt` line directly above them).

No behavior change. These violations went unnoticed because gooddata-eval
is not yet wired into the test/lint harness (see follow-up commit).

jira: trivial
risk: nonprod

gooddata-eval was the only package without a tox.ini, so its tests never
ran via `make test`/CI. As a result the suite silently drifted from the
source as the package was refactored (notably "move all agentic evaluation
logic into gooddata_eval SDK"), accumulating 6 deterministic failures.

Add a tox.ini mirroring the other packages, with `extras = llm-judge` so
the optional `openai` dependency is installed (9 tests import it). Add
pytest-json-report to the test group since the shared pytest command emits
a JSON report.

Update the 6 stale tests to match current, intentional source behavior:
- langfuse test_kind inference now returns "vis_agentic" for visualization
  expected_output
- JSON report run keys are provider-prefixed ("Provider/model") to stay
  collision-free across providers
- ItemReport now requires a `question` field
- run_agentic_visualization only deletes conversations it created; a
  caller-supplied initial_conversation_id is left intact

All 216 eval tests pass via tox.

jira: trivial
risk: nonprod

scripts/docs/python_ref_builder.py imports jinja2, but jinja2 was never
declared in the workspace test group -- it was only present transitively.
Bumping the dev tooling pruned that transitive provider, breaking the
docs-scripts-tests CI job (uv sync --group test --locked; make
test-docs-scripts) with "ModuleNotFoundError: No module named 'jinja2'".

Add jinja2~=3.1 (matching scripts/script-requirements.txt) to the test
group alongside the other scripts/docs dependencies so the synced env
provides it.

jira: trivial
risk: nonprod

parsed_arguments() returns raw model-emitted JSON, so a bad tool call
like {"keywords":[1]} or mixed-type object_types would raise on
.lower()/sorted() and abort the whole eval run instead of scoring
tool_correctness=False. Add _normalize_str_list to drop non-string
entries defensively; valid comparisons (incl. case-insensitive keyword
match) are unchanged.

Addresses CodeRabbit review on PR #1675.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

feat(eval): conversation tracking and --preserve-failed flag

build(deps): bump dev tooling + wire gooddata-eval into tox

Address review nits on search_tool _args_match:
- reword _normalize_str_list comment: dropping non-string entries prevents
  a crash, it does not force a mismatch (surviving strings still compare)
- note that object_types is compared case-sensitively on purpose (controlled
  ObjectType StrEnum values emitted verbatim)

No behavior change. keywords/object_types are declared list[str] in the
search_objects schema, so string-collapse false-negatives cannot occur.

JIRA: TRIVIAL
risk: nonprod

fix(eval): fix search_tool correctness always scoring 0%