The patch tier refused on a whole-file hash mismatch, which is the wrong
question: the three patches touch three small regions of three large files, so
an operator who edited Logging.cs somewhere else entirely was handed a manual
patch job they did not need. Hand-modified shards are the norm, so that refusal
covered most of the audience.
Replace the single hash test with a four-rung ladder (PLAN §2.2.1), cheapest and
safest first:
0 post-patch text already present -> no-op, keeps re-runs idempotent
1 whole file matches the pre-image -> apply verbatim
2 file differs, patched region is still byte-identical -> apply at the
matched offset
3 anything else -> do not touch the file; print the hunk to apply by hand
Rung 2 needs no new metadata: a unified diff already carries the stock text of
the region it edits (context lines plus the '-' lines). Guardrails keep it from
becoming a fuzzy apply -- exact match with only CRLF/trailing-whitespace
normalisation, exactly one occurrence or it fails, line numbers advisory only,
and all-or-nothing per patch file so a half-patched EventSink.cs cannot happen.
install.json records which rung applied each patch, and doctor and uninstall
report it.
This retires the blanket 57.4-only version gate, so PLAN gains §2.2.2 to draw
the line the ladder does not: content matching is a mechanical guarantee about
where text lands, not a support commitment. 57.4 stays the only supported
version. A non-57.4 tree may attempt the tier, but unsupported, untested and not
guaranteed -- behind a loud banner, a prompt defaulted to no, and its own
--patches-unsupported-servuo flag, because a bare --patches can be hit by
accident in a copied script. The unsupported marker persists into install.json,
every later doctor run, and the uninstall report.
INSTALL.md gets the operator-facing half: a block-quoted warning naming the
silent-script-build failure mode, the updated prerequisite row, prompts and flag
table, and a sample run showing all three outcomes.
Co-Authored-By: Claude <noreply@anthropic.com>
