Pre-screen override: land the 16 widened signal patterns, re-measured against the screen_cheap pile

Salvaged from #139, which is closed as stale: the flag machinery it was built around
(`PRESCREEN_ENABLED`, `PRESCREEN_MODE`) was deliberately removed in #161, and the tier
moved out of `run_extract` into Stage 2's `filter/engine/tiers.py`. One part of that PR
never landed and is still worth having.

## What is missing from main

`_SIGNAL_PATTERNS` in `shared/prescreen.py` carries only its original block. The 16
patterns derived in #139 from the vocabulary of 7,505 FLoRA replications and
reproductions — plus the evidence for them — are on `origin/issue-130-prescreen-followup`
and nowhere else:

- the pattern block itself (~30 lines, applies cleanly; `shared/prescreen.py` auto-merges)
- `analysis/prescreen_eval/OVERRIDE_EVAL.md`
- `analysis/prescreen_eval/build_override_sets.py`, `eval_override.py`
- the case sets those scripts read

## Why it matters

`hard_signal()` is a bypass: it is what stops a 3B model getting the last word on a row
whose text states the design outright. A `prescreen_discard` is terminal, so recall here
is not commensurable with dollars. #139's measurement was that the override was catching
79.9% of positives, not the ~100% its derivation set suggested; the 16 patterns took that
to 94.7% (94.1% held out), for roughly $7 a corpus pass in screen calls the tier no
longer avoids.

Two boundaries were tested and declined, and should stay declined: widening
`replication of` to a list of study-like objects (+1.1 points for $6), and the bare
phrase `replication of` (+4 points for $44 and half the screen-confirmed negatives —
"replication of DNA" is the same eleven characters).

## This does NOT widen Stage 1

Worth stating because it is the obvious first question. Stage 1's gate is
`REPLICATION_STEM_PATTERN` in `filter/phrase_detection.py` — a bare stem alternation
(`replicat|replicab|reproduc|reanalys|…`) over the inverted index, unioned with two
OpenAlex concept ids. Every one of the 16 new patterns contains `replicat` or `reproduc`,
so every work they fire on is already in the survivor pool. The override changes which
rows the *cheap tier* may end, nothing about what is discovered.

## The work

The patterns are cheap to apply. The blocker is that the numbers above were measured on
the wrong population for the current design: 7,505 FLoRA papers with usable abstracts
plus 1,333 non-replications "that genuinely reach Stage 3" — from when the pre-screen sat
in front of everything entering Stage 3. Today the tier runs only over rows the rule book
routed to `screen_cheap`, a smaller population selected by rules that did not exist when
the eval ran. Both headline numbers move, and the trade is close enough (the $7 is about
a quarter of what the tier earns; #139 bracketed the tier's live saving at $1–28 a pass,
not the $30 previously quoted) that the wrong denominator could invert it.

So, in order:

1. Re-measure the override's bypass rate over the actual `screen_cheap` pile, and the
marginal screen cost of the rows it newly bypasses. No LLM calls needed for the
bypass rate — it is regex over abstracts.
2. If the trade holds, land the pattern block plus `OVERRIDE_EVAL.md` and the two
scripts, with the re-measured numbers replacing the Stage-3-era ones.
3. If it does not hold, land the subset that does, and record in `OVERRIDE_EVAL.md`
which patterns were dropped and why.

Branch `issue-130-prescreen-followup` is kept, not deleted — it is the only copy of the
patterns and the eval.

Start with shared/prescreen.py and the preserved issue-130-prescreen-followup branch, then read analysis/prescreen_eval/OVERRIDE_EVAL.md, build_override_sets.py, and eval_override.py. Re-measure the 16-pattern bypass rate and marginal screen cost on the current screen_cheap case sets. Done means landing the supported pattern subset and updated evaluation evidence, or documenting dropped patterns and the tradeoff.

Written by the indexing model from the issue text.