docs(readme): bind the field-comparison table to primary sources (B4/L-1)

[Gavin-Borges]

Closes B4 / L-1 - the README "SESTRAV vs Field" table's 36 competitor cells (PredIG / PRIME / NetMHCpan / pVACtools x 9 capability rows) carried no citation of any kind, and had not been checked against their subjects' own sources since the table was written. This is the largest reader-facing integrity gap on the board and a pre-tag blocker.

Every tool's own paper and GitHub README/docs were read directly. Seven cells corrected, twenty-nine confirmed accurate as already written, one deliberately left unresolved.

Corrections

Tool	Row	Was	Now	Source
PredIG	Open source, pip-installable	✓	Open source (GPL-2.0), not pip-installable - Docker/Singularity or webserver	PMC12613480 Code/Data availability; repo README
PredIG	Pan-allele training	Partial	✓ - "PredIG performs pan HLA-I allele predictions"	PMC12613480 Methods
PRIME	Open source, pip-installable	✓	Academic/non-commercial only; precompiled binary or build-from-source	PRIME README, License + Installation
PRIME	Antigen processing as training features	Partial	✗ - its 28-node input layer carries no cleavage/TAP feature	PMC9811684 STAR Methods
PRIME	Pan-allele training	✓	Partial - an "expanded" but still finite allele list	PRIME README v2.1 changelog
NetMHCpan	End-to-end workflow	✗	Partial - auto-digests a submitted FASTA proteome, ranks by %Rank	PMC7319546; DTU 1-Submission.php
pVACtools	Antigen processing as training features	Partial	✗ - NetChop/NetMHCstabPan are optional post-hoc annotation, not training input	pvactools.readthedocs.io
pVACtools	End-to-end workflow	✓ (neoantigens)	same, plus qualifier - the authors' own Abstract says "when paired with a well-established genomics pipeline"	PMC7056579 Abstract

Note the direction on two of these. The PredIG and PRIME open-source/pip corrections overstated a competitor's installability - the opposite of this project's usual error direction. .claude/rules/third-party-claims.md observes that all six prior instances leaned toward flattering SESTRAV; these two did not, which is worth recording rather than smoothing over.

The unresolved cell, and why it stays unresolved

NetMHCpan "Multi-virus support" (Pan-pathogen) is NOT corrected and NOT left quietly as-is. It is flagged in a README footnote and in docs/claims_register.md D34, pending your ruling.

Neither primary source discusses peptide biological origin (viral / bacterial / tumor / self) or uses the term "pan-pathogen" anywhere. The paper's only pan-specificity sentence - "given the pan-specific nature of both methods, predictions can be run for any MHC molecule of known sequence" - supports MHC-molecule breadth, which is a different claim.

Per rule 3, the claim does not get made when the source cannot support it. That cut both ways here: every synthesized replacement considered was itself an inference beyond what either source states, so the right claim could not be sourced either - which is precisely why this is a ruling for you rather than a rewrite by me.

Verification, and two process defects caught

Research ran as a fan-out (one agent per tool, all reading primary sources), then every finding was re-checked by an independent adversarial pass instructed to refute it. That second pass earned its place:

• It caught two citations containing a fabricated or misattributed quotation attributed to NetMHCpan's own paper - a "length distribution of naturally presented peptides" clause and an "any peptide of known sequence" clause, neither of which appears in PMC7319546. This is the direct reason the Multi-virus cell above is unresolved rather than published on the first pass's citation.
• Separately, I caught a transcription bug in the raw research output: one agent recorded pVACtools's "Pan-allele training" current-state as SESTRAV's own column text (No - ten fixed HLA-A/-B binding columns). That is an artifact, not a finding about pVACtools. That row was excluded from this pass entirely rather than corrected from a guess - pVACtools's pan-allele cell is unchanged and explicitly unverified by this pass, and D34 says so.

Both defects are recorded in D34 rather than quietly dropped, because they are exactly the class that file exists to catch.

Harness: the full integrity harness returns byte-identical 134 PASS / 0 WARN / 18 FAIL / 7 SKIP on this branch and on bare origin/main - measured both ways, with and without this diff, so the pre-existing 18 (which #288 addresses separately) are demonstrably not mine. Doc line-citation and commit-reference gates pass; git diff --check clean; zero banned typographic characters in any added line; full fast test suite green via the pre-push gate.