D5: Split phasing-summary.tsv and stratify the switch/flip counts

[TimD1-bot]

Note

Authorship: the content below was drafted by Claude Opus 5 (an AI coding agent) and
filed via gh under @TimD1-bot, a bot account operated by @TimD1. It reflects the
agent's analysis, not a statement authored by @TimD1.

Part of #47. Splits one output file into two and fixes a long-standing format defect in the same
rewrite.

Change

write_phasing_summary (src/phase.cpp:811, declared src/phase.h:50) splits into
write_phasing_variants_summary and write_phasing_blocks_summary.

The split axis is one question: can this quantity be restricted to a subset of variants at all?

New file	STRATUM	Columns	Why
phasing-variants-summary.tsv	yes, first column	VARIANTS, SWITCH_ERRORS, FLIP_ERRORS, SWITCH_ERROR_RATE, FLIP_ERROR_RATE	Each error is attributable to a single query variant, hence to that variant's region sets
phasing-blocks-summary.tsv	no	PHASE_BLOCKS, NG_50, SWITCH_NGC50, SWITCHFLIP_NGC50	Contiguity over genomic span; an NG50 computed over phase blocks intersected with a fragmented subset is not meaningful

The names say per-variant versus per-block rather than "errors" versus "metrics", because #46 adds
non-error columns (phased-variant counts and fractions) to the first file, at which point an
-errors- name would be wrong.

Attribution

ctg_pbs->switches[] and flips[] hold query variant indices (src/phase.cpp:618,625;
confirmed by their comparison against variant indices at src/print.cpp:570,574). An error is
attributed to the region sets of that variant, and switchflips.tsv already reports the same index
in its VARIANT column (src/phase.cpp:751,762,783), so the two files agree by construction.

One convention to document: a switch error is localized to the interval between two consecutive
phased variants, and that interval can straddle a boundary. Attributing to the variant is a choice
made for consistency with the existing VARIANT column, not because the interval has a single
correct answer.

VARIANTS is newly exposed

Today it is only a divisor (src/phase.cpp:824-825), computed as sum(qvars->n) over contigs
(src/phase.cpp:655). Per-subset rates are unauditable without the per-subset denominator, so it
becomes a column.

#46 later substitutes ASSESSED_PAIRS — adjacent phaseable-heterozygote pairs per phase block — for
this denominator. The phaseable subset does not exist until #46 marks unphased variants, so this
issue exposes the denominator it actually divides by, under its true name. Both land before rc1, so
no user sees the intermediate.

The rate-format fix

write_phasing_summary currently emits SWITCH_ERROR_RATE/FLIP_ERROR_RATE with %.6f%%
(src/phase.cpp:826, values already scaled by 100 at :824-825) — a string with a trailing
percent sign in a numeric column — and writes its single data row without a terminating newline.
Both are worked around by every consumer today.

The split rewrites this writer anyway, so it emits bare floats as true fractions —
0.00020930, not 0.020930% — and terminates the row.

The column names do not change, so a consumer comparing a pre- and post-split run reads the same
column with values 100x apart. The legacy % sigil is self-describing, so a reader can discriminate
on the value rather than the column name; that is the mitigation, and the change is release-noted.

Out of scope

switchflips.tsv and phase-blocks.tsv stay unstratified — both are interval/event records rather
than variant records, and a phase block spans many subsets. Deferred, not blocked.

Per-subset NG50/NGC50 is explicitly declined, not deferred: these are contiguity metrics over genomic
span, and computed over phase blocks intersected with a fragmented subset they are not meaningful.

Testing

• phasing-variants-summary.tsv's * row matches the old file's variant-attributed columns, and
  phasing-blocks-summary.tsv matches its contiguity columns — the split demonstrably loses nothing.
• The rate columns parse as bare floats, and the file ends in a newline.
• The current test_phase.cpp contains no write_phasing_summary case, so these writer tests are
  authored here rather than amended.

Depends on

Per-variant membership. Independent of the precision-recall stratum axis.