BFM-decomp/docs/psyq-worklist.md

PsyQ library integration worklist (Phase 8)

Byte-confirmed footprint map of the remaining PsyQ 4.0 libraries linked by SLUS_007.26. Generated by psyq_identify.py over the original EXE (extracted/retail/SLUS_007.26, text 0x80010000..0x800629DC); raw per-object placements in .run/survey/<lib>.txt (regenerable: for L in …; do python3 tools/psyq_identify.py .run/obj40/$L --vram-base 0x8000F800 --exe extracted/retail/SLUS_007.26; done). Replaces the Phase-7 ~350-fn estimate (R14). Placement = candidate footprint; per-object byte-identity is confirmed at integration time by psyq_link_region.py --verify before wiring (the real gate, G3/P9).

DONE (Gen1/Phase 7): libcd (18 objs) + libgs (31 objs) linked byte-identical.

Footprint summary (remaining libraries)

Order	Library	Placed / total	Region(s)	~banked B	Difficulty notes
1	libgpu	3 / 12	800b2 (0x80058890)	~14.9k	✅ DONE (T4 + S78 #4): EXT+PRIM+SYS linked (raw .run/obj40/libgpu, no curated dir). SYS.o was EXCLUDED Phase 8→P31 S78 as scattered-.bss (§9.1); linked since S78 #4 via the link-prepare .bss split (cookbook §489) as libgpu2.
2	libetc	5 / 7	800 tail (0x8004239C)	~3.3k	VSYNC/INTR/INTR_VB/INTR_DMA/VMODE contiguous, ends exactly at libcd1 (0x80043088). Clean.
3	libmcrd	2 / 2	800b2 (0x8005FC68,0x80062888)	~9.0k	LIBMCRD.o (2186 ins, huge, holds the 55 LIBMCRD_OBJ_*) + USERFUNC.o — 2 blocks. _card_* h_norm dups collapse inside LIBMCRD.
4	libc2	17 / 46	800b2 (0x8005C2C8–0x5CD98 + STRCAT 0x80061E90)	~3.1k	C stdlib (BZERO/MEMCPY/STRCMP/PRINTF/PRNT…); contiguous run + 1 outlier. PRNT.o (418 ins) has an internal jtbl — verify NOLOAD .rodata placement (the PRNT_OBJ_24C rodata note).
5+6	libapi+libcard	22 (800c2)	800c2 (0x80061F38–0x80062888)	~0.7k	✅ DONE (combined apicard region): 22 objs / 4 blocks / +24 fns (tools/make_apicard_used.py, C112 dedup, 0 exclusions). libapi's ~22 objects in the 800c3 region (0x5CE18..) DEFERRED — lowest value, separate resegmentation.
7+8	libspu+libsnd	63 / (38+32)	800 sound (0x8003A444–0x8004239C)	~24k	✅ DONE (combined region; S78 #3/#4): 63 objs / 12 blocks. The two libs interleave, so linked as ONE region (tools/make_snd_used.py dedups + excludes 3 addresses; gen_lib_subsegs.py + window). Excluded: S_R/S_W 0x3C438, S_GRMDT* 0x3D424 (cross-object commons — these objects have NO .bss, so the S78 split cannot apply), S_IH/UT_RON 0x3D94C (false-positive, inside SSSTART). VM_F 0x3FA64 (237 ins) rejoined in S78 #4 (snd12, .bss split at _svm_sreg_buf). SSGM.o 0x1BD80 deferred (isolated in matched-C, 8 ins).
9	libgte	58 / 381	800b (0x8004787C–0x5082C) + libgs gaps	~48k	✅ DONE (T11): 53 objs / 22 blocks linked in 800b (subsegs via gen_lib_subsegs.py; integrate window 0x4787C..0x51804). 5 libgs-gap objects DEFERRED (MTX_05/07/11/REG03/REG11 → gsgap1/2/4/5 stay stubs; gsgap2≠MTX_07 exactly so needs sub-split).

SKIP — zero footprint (recorded, not linked by the EXE): libmath 0/48, libc 0/56, libsn 0/51 (2 tiny ambiguous). BFM links libc2, not libc; no libmath/libsn. (Also unbuilt/no-footprint: libcomb, libds, libgun, libsio, libtap, libpress — never converted, no symbols.)

Region map (where each library lives, relative to the current splat subsegs)

800   (0x800123F0–0x80043088)  game code + libsnd(SSGM @1BD80; 3D454–42374) + libspu(3A444–422E8) + libetc(4239C–43088, tail)
libcd (0x80043088–0x8004787C)  DONE
800b  (0x8004787C–0x80051804)  libgte(4787C–5082C, multi-block) + game-code tail
libgs (0x8005080C–0x80057928)  DONE; gaps gsgap1/2/4/5 are actually libgte (MTX_05/07/11, REG03/11; wired S78 #3 as libgte27-30); gsgap3=GS_001 linked S78 #4 as libgs8 (six-piece `.bss` split)
800b2 (0x80057928–0x800629DC)  game code + libgpu(58890–5Bxxx) + libc2(5C2C8–5CD98,+STRCAT 61E90) + libapi(5CE18–626B8 scattered) + libmcrd(5FC68 + 62888) + libcard(61F38–62808)

Integration order rationale (ascending difficulty)

• 1–2 (libgpu, libetc): clean single contiguous blocks, minimal interleaving → re-validate the §9 mechanism end-to-end on a fresh library before the hard ones.
• 3–4 (libmcrd, libc2): few blocks, contained in 800b2; libc2 introduces the PRNT internal-jtbl wrinkle.
• 5–6 (libcard, libapi): short syscall stubs + cross-library aliases + scatter in 800b2 → need windowed placement + curated _used dirs.
• 7–8 (libspu, libsnd): the interleaved sound pair in the 800 subseg (which still holds the 43 REAL game-code matches lower down) → split_src_region.py trims, windowed placement, alias disambiguation.
• 9 (libgte): largest, multi-region (800b + 4 libgs gaps), game-code-interleaved blocks → last.

Per-library integration recipe (each = one byte-gated task, §9.3/§9.5)

Phase 9: all these tools now REQUIRE --vram-base 0x8000F800 --exe extracted/retail/SLUS_007.26 (the EXE's values; psyq_integrate also --symbols config/symbols.us.txt). For a Gen2 binary, pass its own --vram-base/--exe/--symbols (cookbook §9.7, SETUP §6.7).

1. psyq_link_region.py .run/obj40/<lib> <lo> <hi> --vram-base 0x8000F800 --exe extracted/retail/SLUS_007.26 [--emit <p>] → confirm per-object byte-identity ("per-object .text byte-identical: True") + the block/gap structure (G1: derive blocks empirically, don't trust this table's addresses).
2. Curate .run/obj40/<lib>_used (drop cross-library aliases / keep the variant whose .data/.rdata also matches) when aliases exist (make_libgs.sh pattern).
3. Resegment config/splat.us.exe.yaml (one c subseg per contiguous block; non-library gaps stay stubs); split_src_region.py trim the affected game-code .c; make extract.
4. Makefile: add LIB<X>_ELF/OBJDIR/SYMS + a gated psyq_integrate.py --vram-base $(main_VRAM_BASE) --exe $(main_EXE) --symbols $(main_SYMBOLS) … <stubs> [lo hi] call + -T $(LIB<X>_SYMS) on the ld line.
5. Add recovered SDK names to config/symbols.us.txt (mirror to Ghidra, R15/G6); update progress.py's LINKED config.
6. Dual byte-gate: make clean && extract && build && check → 143dbb89… WITH .run/obj40/<lib> present AND with it absent (stub fallback). make report LINKED up. Commit (R20).

Cross-cutting: windowed placement is mandatory for the short-object / interleaved libraries (the libgs GS_106 precedent); cross-library address collisions (C112 libapi/libcard; UT_RON/S_IH; S_R/S_W) mean total distinct objects < the 199 summed placements — byte-verify decides ownership. Any object that won't link byte-identical after §9.1/§9.4 stays a documented honest stub (does not block exit, P9).

⚠️ Boundary gotcha (found at libc2/T6): the game-code subseg boundary AFTER a library block must be the last object's .text SECTION size (8-aligned), NOT its instruction count × 4. psyq_identify reports the instruction count, which omits the object's trailing 8-byte alignment padding (e.g. libc2 SETJMP.o: 30 ins = 0x78, but .text = 0x80 → block ends 8 bytes later). A too-low boundary overlaps the object's padded tail and the relink inserts +N padding, shifting the whole downstream image (pervasive 1-byte reloc diffs + grown file). Verify the boundary = last_obj.vram + readelf .text size for any library whose last object isn't followed object-to-object. Also: when a library block sits at a subseg's START (no < lo prefix to keep), rm the old .c so splat regenerates it under the new boundaries (split_src_region.trim would leave it header-only).

Excluded / deferred objects (§9.1 scattered-.bss) — honest stubs, byte-identical via asm

The standard mechanism places each object's .bss as ONE NOLOAD section at a single base. An object whose .bss commons the original linker scattered to non-contiguous addresses (referenced as .bss+offset by a single section symbol) cannot be reproduced by one base → excluded, kept as an INCLUDE_ASM stub (still byte-identical). Diagnose: psyq_link_region.py <dir> shows N words differ in the object, and its .bss references resolve to >1 base in the EXE. Curate the library's _used dir to drop these.

Object	Library	Evidence	Status
GS_001.o	libgs	scattered .bss (Phase 7) — SIX bases	LINKED S78 #4 (libgs8): psyq_bss_split cuts the section into six single-base pieces at symbol starts (PSDBASEX, CLIP2, PSDBASEY, POSITION, GsDRAWENV); the S77 probe's "5 interleaved bases, NOT splittable" grouped by base instead of by run (cookbook §489)
SYS.o (3109 ins)	libgpu	.bss+0x150 → 0x800c551c but base recovered 0x80078830; commons scattered 0x80078xxx/0x800c5xxx	LINKED S78 #4 (libgpu2): .bss [0,0x144) @0x80078830 + .bss2 = _que @0x800C5510
MTX_05/07/11,REG03,REG11	libgte	sit in libgs gaps gsgap1/2/4/5; gsgap2(48B)≠MTX_07(36B) so the gap stub needs a sub-split	deferred (T11); small GTE fns; link byte-identical, just need the gsgap region resegmented (low priority)
0x3C438,0x3D424,0x3D94C (S_R/S_GRMDT/S_IH)	libspu/snd	cross-object commons referenced at a minority address (S_R/S_GRMDT — these objects have NO .bss of their own, so the S78 split does not apply) + a false placement inside SSSTART (S_IH)	excluded (sound region); stay stubs (24+4+24 ins). VM_F 0x3FA64 (237 ins) left this row in S78 #4 → snd12
SSGM.o	libsnd	isolated @0x1BD80, inside the matched-C region (near func_8001Bxxx)	deferred; 8 ins; would need a 1-object carve amid matched C
libapi 800c3 cluster (~22 objs)	libapi	C57..L10/L02/L03 @0x5CE18.. in the 800c3 region (separate from the 800c2 apicard region)	deferred; ~22 4-ins BIOS syscall stubs; lowest value; another region resegmentation

If scattered-.bss proves prevalent across libgte/libspu/libsnd, escalate to a Max general fix (split each object's .bss into per-common NOLOAD sections at their EXE-resolved addresses); otherwise excluding the few affected objects is the GS_001-precedent decision. Done in P31 S78 #4 — exactly that fix, twenty-three phases later: tools/psyq_bss_split.py runs inside the link-prepare step of psyq_link / psyq_link_region / psyq_integrate and tiles any such section into per-base NOBITS pieces from the bytes (cookbook §489). All three excluded objects link byte-identical; 235 placed objects across the 9 curated dirs, 0 refusals (R39 negative control).

---

The RESIDENT has no linkable PsyQ footprint — it is custom engine code (Phase 12 T1, byte-verified)

Pre-Phase-12 assumption (now corrected): because the resident detects as PsyQ 4.7.0 (DetectPsyQ + a DsMix/libsnd signature hit), it was expected to link embedded 4.7 libsnd/libgte/libspu (the Phase-11 close-out + R24 "link the resident's 4.7 libs" opener). The bytes refuted this.

Phase 12 T1 survey ran psyq_identify over the 4.7 objects (tools/psyq/conv47/…/lib/, already ELF — ar x into .run/obj47/<lib>/, no psyq-obj-parser needed) and the 4.0 objects, against the resident (MAIN.CD/FILE_010/1.1, --vram-base 0x800CEDF8), over the code window and the whole file. Result — NIL footprint:

Lib (4.7)	placed		Lib (4.0)	placed
libsnd	1/226 (ut_rev_2.o, 4 ins)		libsnd	2/163 (VM_DON.o 5 ins, VM_DOFF.o 4 ins)
libspu	0/134		libspu	0/129
libgte	0/509		libgte	0/381
libgpu	0/61
libcd/libmath/libds	1 each (≤8 ins, two alias the same DsMix addr)

All hits are the §9.5 short-object coincidental class (≤8 ins, masked patterns match by luck). Tool sanity (the negative is real): 4.0 libsnd vs the EXE snd region = 35/163 placed — the tool works and is version-sensitive; the resident genuinely embeds no stock PsyQ object of either version.

Ghidra corroboration (G1, sampled): DsMix decompiles to { FUN_800d1bf8(); return 1; } — a custom 2-line wrapper that ignores its vol arg (NOT the stock libsnd DsMix; the R13 tag in symbols.resident.txt is refuted). Other sampled functions are game logic (global accessors, engine init calling EXE REAL matches, entity-heading math calling the EXE's libgte RATAN). The resident's code makes 61 distinct EXE-range jal calls (vs 37 internal) — it calls the EXE's resident SDK rather than embedding it.

Conclusion / architecture: the PsyQ SDK lives in the EXE (959 LINKED); the resident is ~143 functions of custom engine code that calls the EXE's SDK + engine via fixed addresses (no RAM-wasting SDK duplication in an always-loaded blob). The DetectPsyQ "4.7.0" was a single coincidental DsMix-region signature, not a linked footprint. Phase 12 matches the resident engine by hand (REAL), not by linking (LINKED stays 0). R24's per-binary-provenance principle holds, but for the resident the practical consequence is "nothing to link." (Regenerate: for L in libsnd libspu libgte libgpu; do d=.run/obj47/$L; mkdir -p $d; (cd $d && ar x ../../tools/psyq/conv47/psyq-4_7-converted/lib/$L.a); python3 tools/psyq_identify.py $d 0x800CEDFC 0x800D3408 --vram-base 0x800CEDF8 --exe extracted/retail/MAIN.CD.dir/FILE_010.dir/1.1; done)

---

S78 (2026-09-04): the 800c3 "wall" band is LIBPAD 4.2.1 + LIBAPI 4.2, and the LINKED residue is now printed by the build

How it was identified (a free oracle nobody had used). ghidra_psx_ldr ships per-version PsyQ signature sets — ~/ghidra_12.1_PUBLIC/Ghidra/Extensions/ghidra_psx_ldr/data/psyq/<ver>/<LIB>.LIB.json, one masked-byte signature + function labels per OBJECT, for 2.6 → 4.7. Matched against the retail EXE bytes (cookbook §487), the 4.2 set places these byte-exact:

object	vram	ins	functions (labels)
LIBAPI COUNTER.OBJ	0x8005CF68	92	SetRCnt · GetRCnt · StartRCnt · StopRCnt · ResetRCnt
LIBAPI C114.OBJ	0x8005CE48	8	_96_remove — the 4.2 Ps stamp (SETUP §5.1 "libnum 0")
LIBPAD PADENTRY.OBJ	0x8005D0D8	300	PadChkVsync · PadStartCom · PadStopCom · PadChkMtap · PadGetState · PadInfoMode · PadInfoAct · PadInfoComb · PadSetActAlign · PadSetMainMode · PadSetAct
LIBPAD PADMAIN.OBJ (4.2.1)	0x8005D588	~756	PadEnableCom · _padSetVsyncParam · _padChkVsync · _padStartCom · _padStopCom · _padInitSioMode · _padSioRW (+ _padSioRW2/_padClrIntSio0/_padWaitRXready, order-inferred: the 4.2 sig drifts +4/+12 here — the EXE holds the 4.2.1 revision)
LIBAPI L02.OBJ/L03.OBJ	0x8005E168	4+4	SysEnqIntRP · SysDeqIntRP
LIBPAD PADCMD.OBJ	0x8005E188	600	_padSetAct · _padSetCmd · _padSendAtLoadInfo · _padRecvAtLoadInfo · _padGetActSize · _padLoadActInfo · _padSetActAlign · _padSetMainMode · _padCmdParaMode
LIBPAD PADIF.OBJ	0x8005EAE8	~376	statics only (no public labels): func_8005EAE8/EB28/EC00/ECC0/ED4C
LIBPAD PADPORTD.OBJ	0x8005F0C8	408	PadInitDirect
LIBPAD PADSEQD.OBJ	0x8005F728	288	_padInitDirSeq · _dirFailAuto
LIBPAD WAITRC2.OBJ (4.3 sig)	0x8005FBA8	48	setRC2wait · chkRC2wait
LIBAPI FIRST.OBJ	0x80061FA8	168	firstfile (the 4.2 C wrapper; 4.0's A66.o trampoline at 0x80062248 is firstfile2 in 4.2 naming)
LIBAPI PAD.OBJ	0x80062388	192	SetInitPadFlag · ReadInitPadFlag · PAD_init · InitPAD · StartPAD · StopPAD
LIBAPI PATCH.OBJ	0x800626C8	40	EnablePAD · DisablePAD · _patch_pad
LIBAPI CHCLRPAD.OBJ	0x80062768	28	_remove_ChgclrPAD

The 4.2.1x Ps stamp at 0x80072954 ("libnum 12") stands directly in front of libpad's .data — the callback tables D_8007295C..D_800729D8 the band's functions index. So all 12 of main's open stubs in 800c3 (incl. the four §332 "%lo-in-a-delay-slot" walls: _padInitSioMode, _padStartCom, func_8005ED4C, func_8005F450) and the 8 SDK-C-REORDER verbatims are Sony library code assembled in reorder mode — §332b's mechanism, with its provenance. Names applied to config/symbols.us.txt + Ghidra (46; firstfile after retiring Ghidra's 4.0-sig firstfile at 0x80062248 → firstfile2).

Which archives hold it. Placed with psyq_identify after converting each to ELF: PsyQ 4.0 (lib40/, no LIBPAD at all), 4.6 (lib46/, S78 fetch), 4.7 (conv47/): only PDMAIINI.o (4.6/4.7, 68 ins @0x8005D8B4 = _padStartCom+_padStopCom) is byte-identical in the band; 4.6/4.7 libpad otherwise differs (PADMAIN 884 ins in 4.4+, PADCMD 768). The loader's 4.3 signature set matches PADENTRY/PADCMD/PADPORTD/WAITRC2 but not PADSEQD (292 vs 288). A 4.2.1 or 4.3 LIBPAD.LIB is the archive that would LINK the whole band (task #13); until then the band is matched as C under the REORDER_TUS island with the real names.

The LINKED residue, now a build output. psyq_integrate --yaml (S78) wires only the objects inside each library's stub subsegs and PRINTS the rest as ~~ N located object(s) / M ins OUTSIDE the stub subsegs — the completion contract's "SDK residue empties" line, read straight off make build. At S78 close it reads, for libgte: 13 objects / 1,264 ins — MSC01/02/05/09 (800b, 276 ins, 100% of that "game code" subseg), SMP_00 (800b_2), SMP_05 (800b_3), FGO_01–06 (800b_5, 804 ins, 100%), PATCHGTE (800b_6, 40 ins, 100%). Plus, outside libgte's window: SYS.o (800c, 3,109), 2D_BG0/1 (800b_7, 1,022), VM_NO1 (sgap_7, 305), VM_NOWON (sgap_8, 300), VM_F (sgap_6, 237), libapi's 800c3 trampolines (176), FIRST.o (168). Task #3/#4 wire these.

Why the build was RED at HEAD (found S78). The S77 psyq_identify fix (§485) started locating those in-gap objects, and psyq_integrate.contiguous_blocks() then merged libgte's 22 stub blocks into 3 (3 object blocks but 22 stubs) — every make build BINARY=main with .run/obj40/libgte present died at the link. It passed the S77 gates because the gate worktrees have no .run/obj40 and take the stub fallback. Two fixes, both in psyq_integrate.py: (1) --yaml: stub↔objects by SUBSEG RANGE with an exact-tiling check, residue printed; (2) a library object's DEFINED symbol whose recovered address the curated file names differently is --redefine-sym'd to the curated name (R15) — A66.o firstfile→firstfile2, and it also exposed TOC.o CdGetToc@0x800430B8 mis-curated as DecDCToutCallback (an xdedup-vs-VS mislabel; libcd 4.0's linked object is the stronger oracle).

S78 task #3 — thirteen "game code" subsegs were library objects: wired as LINKED (exact tiles, 0 tokens)

new block	was	objects	ins	note
libgte23	800b	MSC01, MSC02, MSC05, MSC09	276	GTE macro wrappers (SetRGBfifo…) that had been hand-matched as REAL
libgte24	800b_2	SMP_00	132
libgte9 (re-derived)	800b_3 + libgte9 + 800b_4	SMP_05 (NormalClip)	12	SMP_06 NormalClipS is its nested sub-pattern; the "3-nop NOTCODE-PAD" was the object's alignment
libgte25	800b_5	FGO_01…FGO_06	804
libgte26	800b_6	PATCHGTE	40	.sbss 0x10, single base
libgte27/28/29	gsgap1/2/4	MTX_05 / MTX_07 / MTX_11	20/12/12	the Phase-8 "deferred (T11)" libgs-gap objects
libgte30	gsgap5	REG03 + REG11	76
libgs7	800b_7	2D_BG0 + 2D_BG1	1022	§485's headline find
snd10	sgap_7	VM_NO1	305
snd11	sgap_8 (head)	VM_NOWON	300	sgap_8 now starts at 0x80040DE8 with its game C

Result: libgte 70 objects / 30 blocks (was 53/22), libgs 33 / 7, sound 62 / 11; LINKED fns 959 → 1,040; REAL 912 → 886 (26 inline-asm SDK wrappers re-provenanced); VERBATIM 146 → 85; 13 TUs deleted; main 143dbb89 byte-identical with and without the SDK objects (the fallback had been red since S7x: CdReadyCallback was called by its SDK name but the libcd stub carried func_800435B4 — curated now). Remaining located-but-unwired: SYS.o 3,109 + VM_F 237 (both .bss-splittable, task #4), the libpad/libapi band pieces (task #5), SSGM.o 8, and the genuine scattered-.bss walls GS_001 / S_R / S_GRMDT.

The metric correction this exposed (R35). progress.py's "MAIN game-code weighted" line claimed its sig excluded the LINKED objects; it never did — the 2026-08-05 Ghidra sig carried all ~31,000 linked-SDK instructions, whose stub records read as unmatched game code. Honest figure, LINKED now excluded live from the Makefile stub lists + yaml ranges: 91.8% (44,562 / 48,537), not 59.8%; the 3,975-ins remainder equals the sum of main's open stubs in frontier_classify exactly.

S78 task #4 (S79, 2026-09-04) — the scattered-.bss class is closed: SYS.o, VM_F.o and GS_001.o linked via a link-prepare section split

new block	was	object	ins	.bss pieces (cut → base)	note
libgpu2	800c (56 hand-matched SDK fns + 62 verbatim frags)	SYS.o	3,109	.bss [0,0x144) → 0x80078830 · .bss2 = _que → 0x800C5510	_que recovered BY NAME from SYS.o's own 4 named refs = 0x800C5510 — the cut is confirmed by an independent oracle
snd12	sgap_6 tail (hand-matched func_8003FA64 = _SsVmFlush)	VM_F.o	237	.bss [0,0x508) → 0x80079580 · .bss2 = _svm_sreg_buf → 0x800B9B58	62 other sound objects recover _svm_sreg_buf = 0x800B9B58
libgs8	gsgap3 (hand-matched as game C: func_800525DC…)	GS_001.o	384	six pieces: 0x80078810 · PSDBASEX 0x800A4F3C · CLIP2 0x800AE820 · PSDBASEY 0x800A4F40 · POSITION 0x800A5E50 · GsDRAWENV 0x800A6438	the S77 probe called this "5 interleaved bases, NOT splittable" — it grouped by BASE; by RUN it is six symbol-aligned pieces, and the other libgs objects recover all five cut symbols at exactly those addresses

Mechanism (tools/psyq_bss_split.py, cookbook §489): a pure-Python ELF32 REL rewrite that derives each reference's base from the game bytes, walks the references in offset order, cuts at symbol starts between runs, moves the symbols, inserts a LOCAL section symbol per piece, retargets the relocations with the addend rewritten in the instruction immediates, and self-diffs. It runs inside the shared link-prepare step (psyq_link.link_object, psyq_link_region.build_region, psyq_integrate.integrate), derived from the bytes on every build — nothing recorded, nothing to go stale. Negative control (R39): 235 placed objects across the 9 curated dirs, 0 refusals, exactly these 3 splits; an end-of-buffer reference (.bss + size, libcd) caught the first bounds check and is why problems are fatal only when a split is actually needed.

Wiring: LIBGPU_ELF is the raw .run/obj40/libgpu (the libgpu_used dir existed only to exclude SYS.o — retired); libgs 34 objects / 8 blocks (make_libgs.sh +GS_001); sound 63 / 12 (make_snd_used.py exclusions 4 → 3). Three TUs went: src/800c.c, src/gsgap3.c, and the _SsVmFlush body of src/sgap_6.c (its 4-ins game function stays). Main 143dbb89 byte-identical WITH and WITHOUT the SDK objects (fallback from a fresh extract). Remaining located-but-unwired SDK code in main: the libpad/libapi band pieces (task #5), SSGM.o 8 ins, and the two cross-object-common walls S_R/S_W + S_GRMDT* (no .bss of their own — a different class from this one).