Drew T
5.1 KiB
§453 ★★★ — ONE SOURCE PER UNIQUE FUNCTION: THE INCLUDE-AT-SITE SHARE, ITS GATE, AND THE FOUR DEFECTS THAT SHAPED IT (P35 S94–S96)
The form. A shared body is a plain-C header under src/shared/<space>/func_<VRAM>[__<h8>|__t<h8>].h (a pure fragment: the
definition, its own externs and comments, no include guard), and each member site is one line, #include "../shared/ov/func_X.h", at
the position the private copy held. The TU keeps its declarations — the body sees the same file-scope environment a private copy saw,
which is why an include is byte-neutral where a separate object is not (Probe P1: three bodies, identical .o, identical relocations).
sotn-decomp's shape (src/st/<name>.h + a per-stage stub), verified from its tree before adoption; the 2026-09-02 belief that "sotn
writes duplicates explicitly" rested on one parenthetical and was false.
The exemplar rule. For an unregistered class the header text is the private copy whose normalized text the MOST copies share (tie →
pin-free → shortest). The same rule applies to a class that is already registered: E_func_80168B70's header carried the 7 late
overlays' spelling (func_80146C3C((u8 *)a0) against their extern void func_80146C3C(u8 *)) and was refused by the 134 main
overlays whose units declare func_80146C3C(void) and call through a cast — same bytes, incompatible environments. Re-exemplar from the
majority (share_body.py --reexemplar): the cast form compiles under both declarations; 141/141 shared, no TU edit.
The gate. Per batch: every touched binary (a twin with its primary) snapshotted, make check BINARY=<b> by exit code, every object
compared; a red binary bisected class by class; the culprit ledgered with the compiler's first real diagnostic. Read the diagnostic from
its FORM, not from the word "error": gcc 2.7.2 prints file:line: too many arguments to function …, conflicting types for …, parse error before … with no "error" token and exits 33 — the first extractor labelled 254 of 303 rejections make: *** … Error 33.
Four defects of the first day, each caught by the run and each a rule now. (1) A tool that restores files with git checkout
restores the COMMITTED text: the second batch's bisect wiped the first batch's uncommitted shares (~50 sites) — restore from an
in-memory snapshot of the batch's pre-edit text, and run one batch per invocation on a committed tree. (2) A twin's instance resolves to
its primary's TU: the same (tu, line) edit queued twice, and the second replacement swallowed the NEXT function — parse error before 'if' and undefined reference to <the following function> in exactly the five twin primaries; dedupe edits by site. (3) The registry
was extended with every planned member BEFORE the gate spoke (788 members, 317 of them private after the run): a listed member is one
whose site shares; extend only with the members that passed, and --repair-registry derives the rest from the census matched on
(binary, vram) — a name-only match removed 9,269 members because h_exact is position-independent and a listed binary can hold the same
bytes privately at another address. (4) add_members_surgical skipped the verbose members: form silently: five 141-member groups sat
at 16 listed; a helper refuses what it cannot edit.
The invariant as a gate (S1). Every same-address class with ≥2 instances is a group with the include at every member, or
twin-covered, or ledgered with its reason. share_census.py --check --strict-macros --strict-text in tools-health, by exit code; C2c
(one source under src/shared/, one definition) and C2d (every member's site includes THAT source) in dedup_integrate --check;
negative-controlled by reverting one include in place (exit 1 naming the class; C2d naming the member).
The second oracle earned its keep on its first run. The sig-blind text oracle (same normalized text, same address, two distinct
TUs) disagreed with the byte join and the disagreement was measured: 3,495 of 4,312 hits were a twin and its primary (one TU), 380 sit
inside the gate-1-deferred cross-address classes (an empty body's hash spans every address), and 38 texts / 2,030 copies were the same
C compiling to DIFFERENT bytes per overlay — the original per-overlay builds compiled that source under different declaration
environments and each of our units carries its own, so no byte tier can class them and no later phase removes them. The h_text tier
(the group hash = the text; every member with its own h_exact; C1 per member) shared all 38 through the same header form, 138/138
green. R34 in one sentence: build the oracle that can disagree, then measure the disagreement before believing either side.
Two published-number corrections the phase surfaced (R14/R75). The fleet denominator 363,221 → 363,680: macro sites invisible to the classifier became includes (ov_SC03_015's 219 single-site macros; the never-extended members of five under-listed groups). REAL 360,744 → 350,533 with EMPTY up the same: empty-bodied shared functions were folded into REAL under the macro form. Instruction-weighted metrics unchanged. A number generated from a form the tooling cannot see is a number, not a count.