Brush's logfile option now writes structured, valid CSV logs that cover the whole run, without any third-party library. This PR also adds a guide notebook showing how to load and plot them.

[gAldeia]

Better run logs

Every call to fit / partial_fit appends to these files. Each row carries a run_id, so several runs can share one logfile.

File	Content
<logfile>	One row per generation: train/validation loss of the best individual, population median/max of loss, size and complexity, plus new run_id, random_state, best_model (the best expression), best_size, best_complexity, stall_count, archive_size, n_evaluations (cumulative)
<logfile>_islands.csv	New. The same statistics for each island, one row per island per generation
<logfile>_simplifications.csv	New. Every replacement the simplifiers actually made: generation, individual id, simplifier (constants / inexact), return type, original → replacement, and distance (MSE between the program's predictions before and after; inexact only)
<logfile>_simplification_table	The dump of the inexact simplifier's stored expressions, now valid CSV and with a run_id column
<logfile>_runs.jsonl	New. One JSON line per run with the seed and the full Parameters

Fixes and improvements

• The simplification table was not valid CSV. Expressions such as Pow(x0,x1) were written without quotes, so pd.read_csv failed on them. All fields are now quoted when needed.
• Headers were repeated. Appending another run wrote a new header line into the middle of the file. The header is now written only when the file is new or empty.
• Rows are flushed after every generation, so the logs are usable during a run and after a crash.
• Individual ids and generation numbers were stale in the variator. It keeps its own copy of Parameters, and its current_gen stayed at the default (1). Ids repeated across generations as a result. It is now synced every generation. Ids are only used for parent tracking and serialization, so evolution results are unchanged.
• Logger::verbosity was uninitialized; it now starts at 0. using namespace std is removed from logger.h.

Implementation

• src/util/csv.h: a small header-only CsvWriter. It quotes fields when needed, writes the header once, flushes every row and writes floats with enough digits to read back exactly. It holds its stream in a shared_ptr, so Engine stays copyable.
• Engine: the statistics loop from calculate_stats() moved into a new summarize(), which is reused for per-island stats. New open_logs() / log_stats() / close_logs(). run_id is a timestamp plus a random suffix from std::random_device, so logging never touches Brush's random generator.
• The simplifiers take an optional SimplificationRecords*. Replacements are only recorded when logfile is set.
• Evaluation::n_evaluations counts calls to assign_fit.

Docs

• New guide notebook docs/guide/logging.ipynb (added to the guide's table of contents). It generates logs for 3 seeds and shows run metadata, convergence, stall count, archive size, evaluations, how the best expression changed, per-island curves, the final Pareto fronts, and replacements per generation with examples.

Tests

• tests/cpp/test_logging.cpp (new): quoting, float formatting, header written once, row width checked, error on different columns.
• tests/cpp/test_brush.cpp: the engine test starts from clean log files and checks the header, one row per generation, and that the other files were created.
• tests/python/test_sklearn_interface.py: fits twice into the same logfile, parses all five files and checks columns, run_ids, per-island rows and that n_evaluations only increases. A new test checks that two runs with the same random_state log identical best_model columns.
• Seeded runs give the same best model with and without logfile.