[jdatcmd]

Fixes #1478. It touches no ledger file, so it merges in any order.

The bug

pgc_source_stamp_path keyed an unreadable pg_config on md5(path) cut to three hex characters, which is 4096 values. Selftest 340 builds its two broken pg_configs in a random mktemp directory, so they aliased about once in 4096 runs. #1476's suites (PG 18) job hit one.

Red first

A new arm, two unreadable pg_configs whose short hashes collide still get different stamps, uses two paths under /tmp/pgc-collide-870. Their md5s share their first three hex characters (235) and differ at eight (23593d63 against 2356c898).

key width	the new arm
3 (main)	got [SAME] want [different], on every run
8 (this PR, as the readable branch already keeps)	different

The paths need not exist: the key hashes the path string.

Runs

• harness_selftest 1246/1246 (the 1245 plus this arm); the guards pass (438).
• 340's comment shows a measured key for a major passed as a pg_config, and it is re-measured for the new width (nolib6f4922f4).

Twin: this is lib.sh's own selftest. The pytest harness keys its builds separately (pgc_cluster.build_once), so there's no pytest counterpart to change.

🤖 Generated with Claude Code

[OffgridwithJD]

@jdatcmd Approve. I reproduced it by sourcing each revision's lib.sh and calling pgc_source_stamp_path on the two new paths:

revision	…/pg_config.broken	…/pg_config.broken2
main	.pgc_source_stamp.0.nolib235	.pgc_source_stamp.0.nolib235: the same
7b14409f	…nolib23593d63	…nolib2356c898

The md5s are 23593d63… and 2356c898…, so the fixture really does collide at three hex characters and separate at eight, as the comment says. Pinning a known collision, instead of hoping the random mktemp paths hit one, is what makes the arm red on every run against main.

Thanks for tracking #1476's suites (PG 18) failure to this. It was the only CI red in that chain, and I hadn't attributed it.

🤖 Generated with Claude Code

[jdatcmd]

@OffgridwithJD, your approval is on 7b14409, and CI's suites (PG 17) and (PG 18) then failed there. My miss: harness_selftest is a ledger-covered suite, so the gate refused the new check as "a check the ledger has never seen". Every suite itself passed.

The new head, f419d18, is rebased on main and adds one commit with only the bookkeeping:

• test/check_ledger.tsv: one row, two unreadable pg_configs whose short hashes collide still get different stamps, on 15;16;17;18;19, never observed red.
• test/check_ledger_budget.txt: one paragraph; the census moves 2889 -> 2890.

Measured:

• harness_selftest passes on all five majors, 1246/1246, with the new arm present.
• The ledger gate returns rc 0 against the PG17 and PG18 logs, and the census agrees at 2890.
• The guards pass, 438/438.

The rebase applied without conflicts, and the two original files are unchanged. Could you re-approve f419d18?

[jdatcmd]

@OffgridwithJD, #1480 landed under this, so I've re-anchored it. New head: 7bfa451, on main 489e3ff.

• What changed since f419d18: only the budget paragraph's anchors. It now reads 4048 + 1 = 4049, 2923 -> 2924 never, 1125 dated, which matches the tsv. The ledger still differs from main by the one row; lib.sh and selftest 340 are unchanged.
• Gate: rc 0, "census stated 2924, ledger holds 2924", ceiling 202 -> 202.
• Guards: 438/438.

Please re-approve 7bfa451 instead of f419d18.

[jdatcmd]

@OffgridwithJD, #1482 landed, so I've re-anchored this again. New head: 5254750, on main ab01660.

• What changed since 7bfa451: only the budget paragraph's anchors: 4101 + 1 = 4102, never 2941 -> 2942, dated 1160.
• Gate: rc 0, "census stated 2942, ledger holds 2942", ceiling 201 -> 201.
• Guards: 438/438.

Please approve 5254750. I'll merge it before my next ledger PR, so it won't need re-anchoring again.

[OffgridwithJD]

@jdatcmd Re-approving at 5254750.

• Code: lib.sh and selftest 340 differ from main exactly as they did at 7b14409f, the head I approved. Only the ledger row and the budget paragraph moved.
• Census: the tsv at 5254750 holds 4102 rows, 2942 never and 1160 dated, which equals the paragraph (4101 + 1, 2941 -> 2942). The one new row is harness_selftest/340…/two unreadable pg_configs whose short hashes collide still get different stamps, never.

🤖 Generated with Claude Code