v0.5.0 — randomHaploid default + --calling-mode

Added

• --calling-mode CLI option (randomHaploid default / randomDiploid /
  majorityCall). Stage 3's genotype-calling mode is now user-selectable instead
  of hardcoded. randomHaploid (the default) and majorityCall are pseudo-haploid
  (0% het) and match the pseudo-haploid AADR 1240K panel — the correct choice when
  projecting a modern WGS target onto ancient DNA. randomDiploid is retained as a
  legacy escape hatch for callers who specifically want diploid genotypes.

Changed

• The .pseudohaploid.json sidecar is now mode-honest. calling_mode records
  the actual --calling-mode used, and pseudohaploid is 1 for the pseudo-haploid
  modes but 0 for randomDiploid (which produces diploid het-bearing calls) — so the
  downstream f2 consumer (pgen-samplebind) never mistakes diploid data for
  pseudo-haploid. Selecting --calling-mode randomDiploid also emits a stderr warning
  (from both the CLI and the orchestrator, so programmatic callers see it too).
• --seed is forwarded for every mode: the random modes use it for read sampling, and majorityCall uses it to break equal-allele-depth ties (so the run stays reproducible).
• validate now probes the installed pileupCaller for all three mode flags.
• Modes are defined once on the CallingMode type (CALLING_MODES,
  PSEUDOHAPLOID_MODES, mode_is_pseudohaploid): the CLI choices and the
  validate probe derive from it, and the pseudo-haploid check is an allowlist
  (a future mode is treated as diploid until explicitly added — fails closed).

Fixed

• Stage 3 now calls pileupCaller --randomHaploid, not --randomDiploid. This is
  a correctness fix. pileup-aadr produces pseudo-haploid genotypes at AADR 1240K sites
  to co-analyze a target with the AADR panel, which is itself pseudo-haploid (one random
  read per site). --randomDiploid samples two reads → a diploid call (~13% het
  on a modern WGS target), yet the .pseudohaploid.json sidecar labelled that output
  "pseudohaploid": 1 "by construction" — mislabeling het-bearing diploid data as
  pseudo-haploid and creating a diploid-vs-pseudo-haploid data-type mismatch with the
  panel in downstream f-statistics (reference-bias + artificial-drift heterogeneity;
  Lazaridis et al. 2017, Souilmi et al. 2022). --randomHaploid (one random read → a
  haploid call, 0% het) is the correct, panel-matching mode. The provenance sidecar
  (calling_mode, note) and Stage-3 status line are updated accordingly. Verified on a
  33× modern WGS target (track_e phase4): randomDiploid → 13% het vs randomHaploid → 0%
  het, with qpAdm admixture weights invariant across calling modes.