KaKsTurbo optional KaKs_Calculator export (introduced in 1.1.0rc9)

Keep result.tsv and results.jsonl as the canonical outputs. Export is a separate,
read-only command on a completed public run; it performs no fitting:
  kaksturbo export completed_run --format kaks-calculator --output new_export
Portable Windows: use .\kaksturbo.exe. Portable Linux: use ./kaksturbo.
Add --json for a machine-readable completion/error record.
Python API: from kaksturbo import export_kaks_calculator
            export_kaks_calculator("completed_run", "new_export")

The new directory contains result.kaks (the exact 22-column header and order of
KaKs_Calculator 3.0), missing_fields.tsv (one reason per unavailable statistical
cell), compatibility.json (provenance, hashes, mappings and limitations), and
README.txt. An existing directory or a location inside the source run is refused.
The source must have a completed public run_manifest/2, all registered final
core artifacts: result.tsv, native_result.tsv and results.jsonl. Their hashes are
verified before and after reading; IDs, order, primary values and counts are
cross-checked. Old public schema-2 outputs are supported. No .work directory,
input alignment, GPU or native computation is needed. Restore the three core
files if moved/compressed; their verification is never bypassed. Other model/QC
artifacts can be separately archived; they are explicitly listed as nonconsumed
and unverified. Export does not recertify the scientific model details.

Robust mode:
Sequence, Method, Ka, Ks and Ka/Ks retain the saved cells, including precision
and NA. The ratio is never recomputed by dividing printed Ka by printed Ks.
Length is 3 * eligible_codons (cleaned nucleotide length). Substitutions is the
observed SNP count. S-Sites/N-Sites and ML-Score retain robust native site counts
and the saved robust MA lnL; their definitions are not certified equivalent to
the baseline estimator's statistics or likelihood.

P-Value(Fisher) is NA: no validated Fisher test is supplied in robust mode.
Fold-site/event statistics, Syn-Subs/Nonsyn-Subs, divergence distance, legacy
rate ratios and GC are NA because their legacy mapping has not been validated.
Expected-event statistics are not relabelled as historical counts. MA-level
AICc/Akaike-Weight/Model are NA: the best candidate and its conditional weight
are not the identity or weight of the complete model average. Candidate values
remain in the original detailed outputs. Other unavailable values remain NA
with their saved reason. No fabricated P value, zero or unit weight is inserted.

Reproduce mode:
Both the 3.0 header and the historical 22-column reproduction header are
recognized. Cells, printed precision and NA are copied unchanged. Five older
labels (S-/N-Substitutions, their Fold fields, and Divergence-Time) map to
3.0 Syn-/Nonsyn-Subs and Divergence-Distance. The ordered Sd/Nd/Si/Vi/t output
fields were checked in both baseline source trees; this is only a documented
label mapping for reproduction mode, not a robust statistic reconstruction. This is still a KaKsTurbo
reproduction-mode result; exporting does not rerun KaKs_Calculator 3.0.

Readers:
  R: source("scripts/read_kaks_compatible.R")
     x <- read_kaks_compatible("new_export/result.kaks")
  Python: add scripts/ to the import path, then
     from read_kaks_compatible import read_kaks_compatible
     rows = list(read_kaks_compatible("new_export/result.kaks"))
Both examples need only the standard R/Python installation. Read tab-separated
fields with quote/comment parsing disabled; keep Sequence as text. The export
does not infer individual gene IDs from the pair ID.

The unmodified GeneFamilies readKaKsCalculatorOutput function was actually run
on exports (pinned commit 777fbfe96edfaddd6c8187b1346c4e23490b9803):
https://github.com/asishallab/GeneFamilies/blob/777fbfe96edfaddd6c8187b1346c4e23490b9803/R/selection_funks.R
It reads Sequence, Ka/Ks and Fisher P values. Its default NA handling converts
a literal Sequence ID "NA" to a missing ID. The supplied safe readers preserve
that ID; the exporter never rewrites it. This test does not certify every
downstream program. A pipeline requiring finite Fisher P values cannot perform
that significance filter on a robust export.

Compatibility means table layout, not estimator equivalence, convergence or
positive-selection significance. Keep the standard results, quality summaries
and model diagnostics, including unresolved projected gradients and NA cases.
The calculation engine, fourteen models, FP64, boundaries, iterations and
correctness thresholds are unchanged. See release QUALIFICATION.json for the
precise installed-package checks. Public downloads: https://github.com/707728642li/KaKsTurbo/releases/tag/v1.1.0rc10
Windows rc10 and Linux rc9-tutorial1 retain this export contract.
