max / makenotwork

Record why the seventeen seed mismatches were not a stale pin

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Author: Max Johnson <me@maxj.phd> · 2026-08-19 20:53 UTC

Signed with PGP, not checked

Commit: 8668f5439b7e42d32ed4d1dfb7eefc370d3a8971

Parent: 2a53c90

1 file changed, +23 insertions, -0 deletions

M server/src/seed/media-manifest.toml +23

			@@ -58,6 +58,29 @@
58	58		# rejected: it would turn "this file was replaced" into a line in a log nobody
59	59		# reads, on a box whose whole job is to be looked at.
60	60		#
	61	+	# STEP 2 IS NOT AUTOMATIC, AND 2026-08-19 IS WHY. Sando build 60 failed
	62	+	# code_smoke with seventeen digest mismatches at once, every one of them a
	63	+	# metmuseum.org URL. Following step 1 cleared all seventeen: the API still
	64	+	# reports isPublicDomain true for each object, each object's title and artist
	65	+	# still match this file, and each `url` is still the object's own
	66	+	# `primaryImage`. Then step 2 refused to apply. Re-fetching from fw13 returned
	67	+	# the digest ALREADY PINNED HERE, for all seventeen, twice over with two HTTP
	68	+	# clients, and six consecutive GETs of one asset were byte-identical with a
	69	+	# stable etag whose `last-modified` is 2023-05-03. Nothing had changed at the
	70	+	# source and there was nothing to repin.
	71	+	#
	72	+	# So a mismatch does NOT imply the pin is stale. It can also mean the fetch got
	73	+	# a variant, and pasting that variant's digest in would pin the wrong bytes and
	74	+	# fail the next build that gets the canonical ones. One oddity worth knowing
	75	+	# while diagnosing: HEAD on that asset reports content-length 1238267 while GET
	76	+	# consistently returns 1222008 bytes, so the origin does describe more than one
	77	+	# encoding of the same object.
	78	+	#
	79	+	# BEFORE REPINNING ANYTHING, re-fetch by hand and confirm the digest you are
	80	+	# about to paste is one you can reproduce twice. If it matches what is already
	81	+	# here, the pin is right and the build is what went wrong: rebuild rather than
	82	+	# edit this file.
	83	+	#
61	84		# Two sources, both chosen because the licence is machine-readable at the
62	85		# source rather than asserted by a human reading a page:
63	86		#