fix: preserve error context across stage / runner / validation paths

[GrigoryEvko]

Four small fixes for the same anti-pattern: error sites that discard useful debugging information before logging or re-raising. All independent, no cross-file coupling.

optimization/utils.py:evaluate_single and the Optuna per-trial path at optimization/optuna/stage.py both rendered ExecRunnerError failures via .stderr.rsplit("\n", 1)[-1] to grab the last line of the worker's captured stderr. That drops the entire traceback above the final RuntimeError: ... summary line — exactly the part a human reading the log needs. Both sites now keep the full stderr in the returned detail string.

optimization/optuna/stage.py:_apply_modifications re-raised the inner SyntaxError as raise ValueError(f"Parameterized code syntax error: {e}") without from e, orphaning __cause__. Logs showed only the wrapped message. Adding from e restores the chain.

ValidateCodeStage re-raised the compile-time SyntaxError as a plain message string — SyntaxError(f"SyntaxError at line {e.lineno}, offset {e.offset}: {e.msg}. Line: ...") — with no .text / .lineno / .offset / .filename set on the new instance. traceback.format_exception then can't render the caret pointing at the offending column. Reconstructing via SyntaxError(msg, (filename, lineno, offset, text)) keeps the structural attributes intact; the original error is still chained. Existing test_syntax_error updated to assert structural preservation (exc.lineno, exc.text) rather than the discarded prefix string.

DagScheduler's timeout and finished-task harvest paths used logger.error("...: {}", e) — only the exception's str() was emitted, no traceback and no __cause__ chain. Replaced with logger.exception(...) so loguru's exception sink renders the full stack.

[GrigoryEvko]

Conflict map vs PR #10 (fix/llm-output-sanitization) and PR #14 (loky-executor)

After the pre-emptive revert on this branch (commit b52544c — dropped the 1-line from e modifier in wrapper.py, which would have been dead code post-PR #14), the conflict footprint is:

vs PR #14: none.

vs PR #10:

File	Hunks	Type
gigaevo/runner/dag_runner.py	11	This PR converts 11 sites from logger.error("...: {}", e) to logger.exception(...); PR #10 wraps the same sites with logger.error("...: {}", sanitize_for_log(str(e)))
gigaevo/programs/stages/optimization/optuna/stage.py	2	ER2 stderr-rsplit reverts + sanitize_for_log overlap
gigaevo/programs/stages/optimization/utils.py	1	Same pattern as optuna stage
gigaevo/programs/stages/validation.py	1	This PR reconstructs SyntaxError preserving structural attrs; PR #10 throws a flattened sanitised string

Resolution (whichever PR merges second):

For the 11 dag_runner.py hunks: logger.exception(...) is strictly more informative — loguru's auto-attached traceback already escapes control bytes — so taking this PR's version supersedes PR #10's sanitize_for_log wrapping there.

For the other 3 files: take the union — keep this PR's structural fixes (full stderr instead of rsplit("\n", 1)[-1]; structurally-rich SyntaxError reconstruction) and apply sanitize_for_log to the interpolated strings. Example for validation.py:

new_err = SyntaxError(
    sanitize_for_log(e.msg or "syntax error"),
    (e.filename, e.lineno, e.offset, e.text),
)
raise new_err from e

The conflict is intrinsic — same lines edited for orthogonal reasons (error-context preservation vs control-byte sanitisation). It is not eliminable at the branch level without gutting most of this PR's scope. Merge order is interchangeable; the second-merged PR needs ~15 line-level unions as documented above.