2026-09-03 - (/loop) The citations were already here, and R9 was already closed

published        348  ->  362
assumed          265  ->  251
differential     211  ->  235 cases
R9                     closed (verified, not done)

Crunching the no-new-capture list. Two of its five items turned out to be smaller than stated, and one of those corrections is mine to make.

Fourteen entries were bookkeeping, not research

math.rs adds ten float functions in one block from one standard. The first line names it:

// `acosf(x)` - ISO/IEC 9899 7.12.4.1.
// `asinf(x)` - 7.12.4.2.

acosf is published with that citation. asinf is assumed, with 7.12.4.2. sitting in its purpose field and the note "cites no published specification" - which was false. One block, one generator, six characters of difference.

Fixed in the comments rather than in the generator: teaching it to accept a bare 7.12.4.3. would be the provenance-manufacturing its own guard exists to prevent. Twelve comments now name the standard - better documentation anyway, since a reader of one line should not have to scroll to a header. strtoimax/strtoumax said C99, a name the list does not carry; spelled ISO/IEC 9899 they cite their exact clause.

Verified by deleting all fourteen and letting the generator rewrite them. All fourteen came back published; thirteen with the exact sub-clause. hypotf came back with only the bare standard name because prose followed the clause on its line, so that line was split and it regenerated exact. D501.

And the correction I owe: the rest is not citation work

I said a large share of the 265 had a documented analogue waiting. That was wrong. Ninety- nine are libScePosix, and their assumption reads:

That the POSIX name and scePthreadMutexLock are the same behaviour rather than merely similar. Unmeasured.

No standard settles that. POSIX says what pthread_mutex_lock does and nothing about the vendor's export of that name - and D468 is this project already paying for that exact confusion once, with the ctype tables. Relabelling them would have been that mistake ninety-nine times.

The remainder are Dinkumware CRT internals, Itanium-ABI throw helpers, and vendor libkernel functions with no analogue. The assumed count is close to honest.

R9 was closed, and nobody had checked

provenance_faults carries learn's rule and is wired into check(); a test asserts it empty on every run. The second half - every implementation has a knowledge entry - is also done: 694 entries against 475 bound implementations, and the five names my first query flagged as missing were test fixtures (abi, loader, mem, thunk, worker), not guest functions.

Fifth stale work item this week, after /dev/random, the differential's "missing" cases, D488's note on the verdict, and Phase 0d.

Also measured, not acted on

The 95 outstanding hardware measurements are not mostly actionable. Thirty-nine of them are 106-encoder/* - "whether a symbol resolves inside a video-encoder library" - and orbistoun implements no encoder subsystem, which principle 6 puts well after the address space and threads. They are correctly parked, not neglected.

And the differential covered everything except the two most-called functions

211 -> 235 cases. Checking the case list properly this time rather than trusting the plan's list, the gap was not exotic: strlen and memcpy had no case between them, and they are the fifth and sixth most-called functions in the whole recorded corpus at 89,811 and 97,664 calls. A differential covering strtoumax and not memcpy is measuring the wrong end.

Added with abs/labs/llabs and atoi/atol/atoll - the latter being strtol with the error reporting removed, which is exactly why they need their own cases rather than being assumed to follow it. INT_MIN is deliberately absent from the absolute values: negating it is undefined, and a case for it would record one implementation's choice as the contract.

All 235 agree. Watched failing by corrupting one byte of the memcpy window, after the run_scalar refactor changed the dispatch again.

Three of my own mistakes, caught by reading the output rather than by the gate:

• A shell heredoc collapsed nine escapes into real newlines, so the reference did not compile - and the redirect had already truncated the committed file to zero cases. Write to a temp, verify, then install; the script now refuses to install a short run.
• "a€b" is not a high byte followed by b. A hex escape in C has no length limit, so that is the single character 0x80b truncated - strlen measured 2. Split the literal.
• Before that, the same escape had been typed as a character and arrived as two bytes of UTF-8, measuring 4. The case is named "embedded-high-byte" and was about neither.

State

cargo test --workspace green - 117 suites, 1992 tests, 0 failures. clippy --tests clean, fmt clean, identity scan clean.

Nothing committed. The day holds worklogs 292-351 and D466-D501.

Scoreboard for the no-new-capture list

Five items were named. Measured, three of them are smaller or already closed:

Next: the differential is the one item with headroom left - strtok_r, strtof, sprintf/vsnprintf and the wide-character functions are uncovered, the last needing a wide text encoding in the record format.