[Jammy2211]

The fix in #63 was aimed at the wrong half

#63 taught normalise_title to reconcile $0.5\leq z<1.0$ with 0.5≤z<1.0.
The dry run that followed resolved 19 of 74 — none of them the maths titles it
was written for. Comparison was improved when retrieval was what failed.

resolve_title built its phrase by replacing every non-word character with a
space, so a title carrying maths went out flattened word-by-word:

ti:"Photometric properties of … galaxies at 0 5 z 1 0"
ti:"Dwarf galaxies imply dark matter is heavier than 2 2 10 21eV"
ti:"Review of solutions to the Cusp core problem of the ΛCDM Model"

arXiv indexes the LaTeX source and matches none of that, so no candidates came
back — and a better comparison has nothing to compare. (Λ survived into the
query verbatim, because Python's \w counts Greek as a word character.)

The change

query_phrase drops the maths rather than mangling it, searching on the longest
run of ordinary words the title contains — the part both sides spell the same
way:

title	query sent
…galaxies at 0.5≤z<1.0	ti:"…galaxies at"
…heavier than 2.2×10−21eV	ti:"Dwarf galaxies imply dark matter is heavier than"
…problem of the ΛCDM Model	ti:"…problem of the CDM Model"

This is a broader search, not a looser match. The phrase only decides which
candidates come back; every one still has to survive the exact, unique-title
check in match_entries, and an ambiguous set still resolves to nothing. A test
pins that distinction directly — a candidate sharing the searched phrase but not
the full title is rejected. max_results rises 20 → 50 to suit the wider net,
which costs nothing since the choice is made locally.

Workflow, so the next pass is usable and watchable

• mark_unresolved dispatch input, wired to --mark-unresolved (the marker
  added in queue: line kinds — a heading is not a paper you can download #62). Only reachable from a deliberate dispatch: inputs.* is empty
  on a cron fire, so the nightly can never mark on its own.
• python -u — the log streams during the run instead of arriving in one
  block at the end. This is why every run so far has been unreadable while in
  flight.

5 new tests; 121 pass, make validate clean.

---

Generated by Claude Code