sim report: word an empty-side Avg win/loss (R) by counting the R sample (#1004)

[eaitbrahim]

Closes #1004

What

The keel sim edge table's Avg win (R) / Avg loss (R) cells printed the engine's empty-side 0 as 0.000, which reads as a measured flat side. They now print:

R sample	Avg win (R)	Avg loss (R)
all wins	measured	n/a (no losses yet)
all losses	n/a (no wins yet)	measured
all scratches	n/a (no wins yet)	n/a (no losses yet)
a side whose trades average exactly 0 R	0.000	0.000
no R at all	n/a	n/a

How

• keel/commands/insights.py: new avg_r_reason_of(stats, outcome), beside profit_factor_r_reason_of. It decides the side by counting the R sample (n_wins_r / n_losses_r), never by reading the average, and reuses the existing NO_WINS_YET / NO_LOSSES_YET / NO_TRADES_YET constants. No new strings.
• keel/sim/report.py: new _avg_r_cell, built the same way as _profit_factor_r_cell: bare n/a with no R, then undefined_reading(reason), otherwise _r_cell(value). keel.commands.insights is still imported lazily, so the existing subprocess import-isolation test keeps passing.
• BacktestResult, stats._r_aggregates, promotion.pool_stats and the gates are unchanged. pool_stats still multiplies by the engine's 0.

Tests (TDD)

• test_edge_section_words_an_empty_side_average_r_by_counting_the_r_sample fails on main (('wins:A', '1.500', '0.000')) and passes here. It covers all-wins, all-scratch, all-losses, a hand-built zero-R win plus zero-R loss (to check that sides are counted, not read), both sides, no trades, and the pooled row.
• test_avg_r_reason_of_an_empty_r_sample_is_no_trades_yet_on_either_side covers the empty-sample and all-unrisked branch.
• Mutation-checked 7 mutants, each confirmed applied with cmp and each killed: reading the average instead of the count (both sides), a swapped count, dropping the empty-sample branch, dropping the no-R guard, reverting the loss cell, and an outcome-blind value.
• Full suite: 8461 passed, 3 skipped. ruff check, ruff format --check and bare mypy are clean.

🤖 Generated with Claude Code

[eaitbrahim]

Review of #1007. The counting-over-reading logic, the bare n/a for no-R rows, the lazy import, and the unchanged engine values / pool_stats all check out. The tests pin values per row, and the fixtures are asserted to reach each case. The 0.000 zero-R sides and the unrisked-sample no trades yet case would catch the obvious mutants. One false docstring claim, inline.