feat(citations): version groups (merge preprint + published citations)

[neuromechanist]

Problem

OpenAlex keeps separate work records for a paper's preprint and published versions and splits citations across them. Tracking only the published DOI undercounts badly:

• LSL: published (Imaging Neuroscience) = 61 citations, but its bioRxiv preprint holds 98.
• BIDS Apps: the published PLoS Comp Biol record shows 0 (OpenAlex lost them); the preprint has 56.

Fix

A canonical paper can declare version aliases. The citation sync resolves every version DOI to an OpenAlex work id and queries them together with an OR-joined cites:W1|W2 filter, which OpenAlex deduplicates, attributing the merged per-year histogram to the primary DOI.

citations:
  aliases:
    "10.1162/IMAG.a.136":            # LSL published
      - "10.1101/2024.02.13.580071"  # LSL bioRxiv preprint

• CitationConfig.aliases (primary DOI -> version DOIs), validated/normalized like dois.
• OpenAlexCitationClient.counts_by_year / recent_citing_papers accept a work-id group; _cites_filter OR-joins with |.
• sync_citing_papers(aliases=...) builds the group from primary + aliases; CLI and scheduler pass the community's aliases.
• Configs: eeglab LSL preprint, bids BIDS Apps preprint.

Verified against OpenAlex: LSL combined per-year rises 61 -> 157 (2025: 75, 2024: 34, ...).

Test plan

• Config: aliases default empty, primary+version normalization + dedup, invalid-DOI rejection.
• OpenAlex client: _cites_filter single/multi/empty; counts_by_year sends the OR-joined filter for a group.
• End-to-end: sync_citing_papers with aliases resolves both versions, OR-joins (cites:W1|W2), and attributes the merged count to the primary DOI (mock transport, real DB).
• Regression: config, citations endpoint, cli sync, scheduler — all green (231 passed).

Deploy follow-up

Re-run sync papers --community {eeglab,bids} --citations so LSL/BIDS Apps pick up their preprint citations.