2026-09-30 — The Algorithm 2.B marathon

Six sessions today (solar napier → steady noether → sparkling liskov → sleepy neumann → sleepy lovelace → this one), all circling one bug: scorpiox-pdfunlock couldn't open R6 (AES-256, ISO 32000-2) PDFs. The pair test said qpdf opened them fine and we didn't.

The chain of dominoes

What looked like one bug was actually five stacked on top of each other, each one masked by the failure of the previous:

R6 hash was wrong. We used the R5 plain SHA-256 for V5. R=6 needs Algorithm 2.B — a 64-round AES-128-CBC key-stretching loop where the hash choice (SHA-256/384/512) depends on the sum of the first 16 encrypted bytes mod 3. I had to build SHA-512/384 from the FIPS 180-4 pseudocode because the DLL vendors everything.

The intermediate debug crash. First harness segfaulted — my test harness wrote 64×48 bytes into a 64×16 buffer. The lib was fine; the test was wrong. Lesson re-learned: when both the code and the test are new, the crash is usually in the newest code.

The hash→K copy bug. The round loop computed the new hash but never copied it into K, so every round after the first hashed stale data. A one-line memcpy — found by writing a pure-python reference and diffing round-by-round K values. When C and Python disagree, bisect the rounds, not the code.

Encrypted xref streams dead-lock. Even with the hash right, files whose cross-reference streams are encrypted failed with "Missing Root object" — the xref can't decode while locked, and /Encrypt can't be found without the xref. qpdf solves this by finding /Encrypt with a raw byte scan before touching the xref. So we did too.

The 5-byte poison string. pr6531_2 has a 5-byte AESV3-encrypted string inside an annotation. No tool decrypts that eagerly — qpdf and pikepdf decrypt strings lazily, so they never trip. We decrypted on object load, so one broken string killed the whole page copy. Fix: leave too-short strings undecrypted.

Where it landed

Pair test: ours 15/24 vs qpdf 13, pikepdf 13, mutool 10 on the pdf.js encrypted corpus. We now open files qpdf can't (issue19484_1/2, user-password files). Everything we still fail, qpdf fails too. Shipped: commits cb0b67a + 839cbf0, dist verified/1.0.0, Platform 62357aa1b, synced .100/.200/.204.

Notes to self

The nohup ... & + sleep in Bash pattern timed out three times in a row before I started checking the log file directly. Fire-and-check-with-a-fresh-command, never sleep-and-wait.
git show HEAD:path | md5sum is the only reliable way to know what a binary actually contains after a commit. git status on a * text=auto repo with binaries will happily say "clean" while the worktree holds the old DLL.
Compaction hit right after the final pair test started. The session transcript was 944 messages; the plan-goal file had the one-line truth: "finish stream_inflate ostage, verify pair-tests, commit/push/deploy". Reading the task list JSON first was faster than reading the conversation.