consensus: fix six gene-structure bugs reported in #4

[xiekunwhy]

This PR fixes the six consensus bugs reported in issue #4, plus a sixth one we
found while validating the fixes (fully-masked introns splitting real genes).
Each fix carries a targeted unit test; the full suite passes and the golden
consensus output on the EVM testing data set is unchanged.

1. partial3 semantics — a mid-gene premature stop (the ORF projection
   truncates the CDS and demotes downstream path exons to three_prime_UTR)
   was reported as a complete 3' end. partial3=false now requires the
   projected CDS to reach the terminal exon (empty three_prime_UTR).
   In our plant-genome run, 4,456 of 36,439 screened "complete" genes were
   silently truncated this way.
2. GFF3 phase column honored — prediction CDS rows' column-8 phase was
   discarded and every chain assumed to start in frame 1, frame-shifting all
   5'-partial input predictions (Helixer/Tiberius/AUGUSTUS emit these
   routinely). Each chain now carries the phase of its 5'-most CDS row
   (translation order) and uses it to seed inherited frames.
3. Stop-screening escape hatch — try_add_exon inserted exons even when
   determine_good_phases found an in-frame stop in all three frames
   (the guard only fired on non-empty results).
4. Conditional 3 bp trim — terminal/single exons now trim their trailing
   3 bp only when those 3 bp actually encode a stop under the configured
   genetic code; pipelines whose CDS rows exclude the stop codon (separate
   stop_codon features) no longer lose 3 genuine coding bases.
5. No unchecked-complete fallback — the genome-fetch-failure fallback in
   resolve() emitted the raw path exons as CDS with no stop scan and could
   mark it complete; hit_stop is now false there.
6. Fully-masked introns registered with score 0 (EVM parity) — EVM
   registers an intron even when repeat-masking zeroes its score
   ($INTRONS_TO_SCORE{$key} += 0), keeping trellis connectivity; combinr
   skipped registration entirely. In a genome that is 80% repeat-masked this
   split real genes whose introns are covered by TE insertions (one concrete
   BUSCO-complete-in-EVM locus dissected in Some issues #4 follow-up). Masked bases still
   contribute no score, so the anti-TE purpose of --repeats is preserved.

Validation: cargo test all green (including the new tests for each fix);
A/B against the upstream binary on the EVM testing/ set is byte-identical.

A follow-up PR (see follow-up PR for the flags) builds three opt-in production flags on top of this one.