ROM 1.3: the source disk, and what is left
The machine · partially solved
Status: Partially solved. The disk is read
and the source is extracted, indexed and readable. What is verified: the
DBE/DCE rule, and a rev-B control build that reproduces a shipped Osborne ROM
to 3759 of its 4096 bytes. What is not: the 1.3 image itself, which is
reconstructible but not certified, and must not be handed upstream as the real
ROM. Two things stand in the way and both are tooling, not knowledge: the 1.3
source uses the DBE/DCE pseudo-ops, which only ACT 3.5E has, and every
build loses 337 bytes of its hex output to an emulation defect (§7). Everything
needed to finish it is in this repo or named below, and the recommendation is to
record the result and stop (§10). Issues: #205, #72, #70, #203.
Vocabulary:
rev Bhere is the Osborne 1’s monitor ROM (9/12/81) — a 1979-era ROM part, not a board revision. Board revisions belong to the boards; see [hardware/README.md](https://github.com/Myrcurial/o1next/blob/main/hardware/README.md).
MAME carries BIOS A, 1.2, 1.2.1 and 1.3 as NO_DUMP placeholders, so
mame osborne1 -bios ver13 cannot run. We hold the 1.3 source disk —
research/source-code/OSROM13.IMD, whose own comment reads
rom 1.3, cbios .3 osborne source 12/16/81 — and this is what came of
trying to turn it into the missing osb13.bin.
The answer, up front
- The disk reads, and the source is complete.
OSROM13.IMDis an 8-inch disk, not an Osborne 1 5.25” one, which is why no earlier reader in the tree could describe it. It yields 19 files —OCCROM0..9.ASM(the monitor ROM, split into modules) andOCCBIO0..8.ASM(the CBIOS) — all of them clean text that starts with aTITLEand ends with; END OCCROMn?.???. They are committed atresearch/source-code/ROM-13-source/. - The build needs ACT 3.5E, and we do not have it. The source uses two
pseudo-ops,
DBEandDCE, that exist in no ACT the archive holds — everyACT80.COMin the local archive is 3.0 or 3.2 (or ACT II 8086 3.4), and none of them carries those strings in its mnemonic table. The rev-B listing onOCCROM.IMDnames the version it was built with:SORCIM 808x Assembler ver 3.5E. - What
DBE/DCEdo is recovered exactly, from a shipped ROM we do have: the rev-B disk carriesOCCROM0B.COM(4096 B) and its full listing, and comparing the two gives one rule holding over every line. See below. - A 1.3 image can be produced —
tools/research/rom13-reassemble.shdoes it, by rewritingDBE/DCEintoDBso ACT 3.0 emits the same bytes — but it is not certified: three errors in the 1.3 build are still unidentified, and 337 bytes of every build’s hex output are lost before they reach the disk. It must not be handed upstream as the real 1.3 until the control build passes. §6. - The hex output arrives damaged at about six fixed file offsets per build — identically on a double-density disk and on a single-density one, and on drive A or drive B — so it is not the disk layer. §7.
- Recommendation: stop here, as “record it and stop”. The source work is
done and committed; what is left is two tooling defects (§6, §7) whose
clean fix is the assembler we cannot get. Nothing in the project needs a 1.3
ROM, and the durable half of the result — the disk read, the source, the
recovered
DBE/DCEsemantics, and a control build that reproduces 3759 of a shipped Osborne ROM’s 4096 bytes exactly — is already banked. §10.
1. What the disk is
imdinfo.py reports the geometry; tools/research/o1imd8.py reads it.
| Cylinders | 77, one head, 630 784 bytes |
| Track 0 | 250 kbit/s FM, 26 sectors of 128 bytes (the IBM-3740 boot track) |
| Tracks 1–76 | 250 kbit/s MFM, 8 sectors of 1024 bytes |
| Filesystem | CP/M 2.2, 2 KiB blocks, 128 directory entries (4096 B = two blocks), two reserved tracks |
| Skew | 3 — the sectors of a track are interleaved |
Seven disks in research/source-code/ share this format (OCCCBIOS, OCCROM,
OCCUTIL, OSBIOSSR, OSBROM, OSROM13, OSUTLSRC), so the reader is a
tool, not a one-off.
The skew is the whole difficulty
o1imd8.py writes the image in the linear shape a cpmtools diskdef expects
(track 0’s 3328 bytes padded out to 8192, boottrk 2), and the diskdef carries
skew 3. Without the skew the read still works — the directory parses and
every file comes back — from the wrong sectors. That is the failure mode
docs/ozrom-1e.md §10.4 warns about, and it is how this one was found:
content chaining. At skew 3 every sector boundary in occrom33.asm joins
mid-token to the next —
BDOS (f + unction -> BDOS (function
\tJMP\t + LIST -> \tJMP\tLIST
;HOUSE + KEEPING -> ;HOUSEKEEPING
\tLD\tA,PIAB + D -> \tLD\tA,PIABD
— and at skew 1 the joins are ;continueT, indicate RY and a sector that
should not follow. skew 3 is also exactly tools/cpmfs.py’s
skew_table(3, 8) = 0 3 6 1 4 7 2 5, i.e. cpmtools’ own table
(cpmfs.c), which is the cross-check that the two agree.
2. The source set, and what is missing from it
OCCROM03.ASM is the master. It sets ORG 0, defines ALPUSH/ALLPOP, and
LINKs the rest:
LINK OCCROM13.ASM ;ROM STANDARDS
LINK OCCROM23.ASM ;CPM BOOT
LINK OCCROM33.ASM ;CONSOL ROUTINES
LINK OCCROM43.ASM ;PIA ROUTINES
LINK OCCROM53.ASM ;SIA ROUTINES
LINK OCCROM63.ASM ;DISK ROUTINES
LINK OCCROM73.ASM ;DIAGNOSIC ROUTINES
LINK OCCROM83.ASM ;FORMAT ROUTINE
LINK OCCROM93.ASM ;UTILITES
LINK OCCRAM2.ASM ;RAM LOCATIONS
| File | Size | On OSROM13.IMD? |
|---|---|---|
occrom03.asm … occrom93.asm (10) |
896 … 32 512 B | yes |
occbio03.asm … occbio83.asm (9) |
640 … 16 896 B | yes |
occram2.asm |
3840 B | no — recovered from OSUTLSRC.IMD |
occtxt6.ast |
9472 B | no — recovered from OSUTLSRC.IMD |
Two files the build needs are not on the ROM disk at all. OCCRAM2.ASM is the
;RAM LOCATIONS module (ORG MRAMEXT, ORG MRAM); the utilities-source disk
OSUTLSRC.IMD carries it. The systext (OCCTXT) is the equates file the ROM
reads for CR, LF, MTRK, EOFC, MRAM, MRAMEXT, KL_LEN, ALVS,
CSVS and the rest; no OCCTXT is on the 1.3 disk either, and the one
committed here is occtxt6.ast from OSUTLSRC.IMD — the nearest revision we
hold to 12/16/81. That is an assumption, and it is the one place a future
find could change the bytes: occtxt5.ast and occtxt6.ast are identical
apart from the header and BIOS, and BIOS is referenced by no ROM source
line, so the choice is inert for this source — but a later OCCTXT is not
guaranteed to be.
3. Obstacle one: CCP is defined twice
The vendor’s recipe for the CBIOS is ACT OCCBIO05.ASM S=OCCTXT6.AST R=F L=A:
(bios141.gen), so S= is how the systext is fed in. Feeding occtxt6.ast to
the ROM master fails immediately and fatally:
*Error page 0 SORCIM 808x Assembler ver 3.0 00/00/00 00:00 Page 1
Osborne Computer Corp System Text. A:OCCTXT6 .AST
D = D000 CCP = BIOS-1600h ;CCP entry
1 ERRORs, 716 Labels, 75A7h bytes not used. Program LWA = 0000h.
The master sets CCP = 0 (“Developed for use with the new cbios”) and every
OCCTXT we hold defines CCP = BIOS-1600h. ACT 3.0 treats the duplicate as
fatal and abandons the assembly. The 1.44-era systext (SVER = 14) has
removed MSIZE/BIOS/BDOS/CCP altogether, which is evidence that this
line was on its way out; the revision contemporary with 12/16/81 is not in the
archive.
rom13-reassemble.sh therefore disables that one line in its copy of the
systext, leaving the source alone. CCP is not referenced by any ROM source
line (it appears only in the master’s own definition and in comments), so the
change cannot move a byte of object code — and the rev-B control build, whose
master does not define CCP, needs no such change and passes the same way.
The edit lives in the script rather than in a source file, and the tooling
hazards that surround the build are in §7.
4. Obstacle two: DBE/DCE are ACT 3.5E only
Assembling without a systext gives 241 errors, all U (undefined symbol)
on things like CMP MTRK, CPI EOFC, LK C,CR, LDK HL,D.TRKR — the systext
fixes those. With the systext and the CCP line disabled the count drops to
27, and every one of them is on a DBE or DCE line:
I 0E3B 000000 :xmsg: DBE cr,lf,'Program begins' ! DCE ':'
Neither mnemonic exists in ACT 3.0 or 3.2 — not in DOCOP80.ASM (which lists
all 130 ACT mnemonics) and not in the mnemonic table inside ACT80.COM, which
is plain text: … DA DB DCR DCX DC DEC DI DJNZ DS DW …, with nothing between
DB and DCR.
What they emit, recovered from a shipped ROM
OCCROM.IMD (the rev-B ROM source, 9/12/81) is the disk that makes this
answerable: it carries the source (which uses DBE/DCE), the shipped
image OCCROM0B.COM (4096 B), and the full listing OCCROM0B.PRN — so
every DBE/DCE line can be read against the bytes it produced. The listing gives
the address of each line and the ROM gives the bytes:
0D3A 34 DIAMSG: DBE 'Z'-40h
0D3B 404040A49E DBE ' ROM DIAGNOSTICs'
0D4D 1A14A6CAD8 DBE cr,lf,'Select Test:'
0D9E A6CAD8CAC6 DCE 'Select (A,B): '
Decoding those (and the other eleven lines) gives one rule, with no exceptions:
byte = value * 2 DBE — every byte even
byte = value * 2 + 1 DCE — the *last* byte only
0D3B: ' '=0x20→0x40 ×3, 'R'=0x52→0xA4, 'O'=0x4F→0x9E, … 's'=0x73→0xE6.
0D4D: cr=0x0D→0x1A, lf=0x0A→0x14, 'S'=0x53→0xA6, … ':'=0x3A→0x74.
0D9E (a DCE) is the same shift, and its last byte is 0x41 = 0x20×2+1 —
the end-of-string marker. The rev-B ROM contains 18 literal 0x7F bytes, but
none of them falls inside a DBE/DCE string, so the “7F, repeat count, char”
escape rom144.asm documents is not exercised here.
tools/research/actdbe.py rewrites each DBE/DCE statement into DB
statements whose operands are the same expressions multiplied by two, leaving
ACT to evaluate cr, lf and the rest itself, and keeping the original
statement as a trailing comment so the result is auditable. Two things the
translation has to get right, and both fail silently: the CP/M source is CRLF
(emit bare LF and the source text lands in the hex file as object code), and
! is ACT’s statement separator while ACT’s local labels are :name:
with a leading colon.
5. What is verified, and what is not
| Claim | Evidence |
|---|---|
| The disk reads and the source is complete | 19 files, sizes matching the directory, all clean text, all ; END OCCROMn?.??? terminators |
| The format is skew 3 | every sector boundary in occrom33.asm chains mid-token at skew 3 and nowhere else; matches cpmfs.py/cpmtools’ own table |
DBE = value×2, DCE = value×2 with the last byte +1 |
15/15 lines in the rev-B listing against the shipped OCCROM0B.COM |
| The translated rev-B build reproduces the shipped rev-B ROM, as far as its output survives | no ERRORs; LENGTH OF THIS ROM IS = 0FC6, the shipped listing’s own number; and 0 differing records against the shipped hex — 3759 of the shipped image’s 4096 bytes byte-identical. The other 337 bytes are the 16 records lost to §7, so they are unverified, not verified |
| The translated 1.3 build is the ROM | NO — 3 errors remain and 337 bytes are lost to §7; see §6 |
| The built 1.3 image boots | NO — run, and it fails. MAME renders it as a full screen of block glyphs, no banner. Method and screenshot: §6.1 |
| The vendor-built rev-B image boots | YES — OCCROM0B.COM (CRC32 f5f697c7) prints OSBORNE SYSTEM ONE.M / REV B 227 and INSERT DISK IN DRIVE ... AND PRESS RETURN. under MAME. §6.1 |
6. The control build, and what still stands in the way
The check that matters is the rev-B control build: build rev B from
OCCROM.IMD with the same translated source and the same ACT 3.0, and compare
against the OCCROM0B.COM that shipped on that very disk.
The assembler half of that passes. One actdbe.py defect had to be fixed
first: an unmatched optional group in the label regex emitted the literal text
None as a label on every unlabelled DBE/DCE statement, so ACT defined a
symbol called None over and over — 25 errors in the 1.3 build, and in the
rev-B build an LWA of 0FAAh against the shipped listing’s 0FC6h, with
labels late in the ROM resolving 28 bytes low. With the regex fixed
(label or ''), rev B gives:
'LENGTH OF THIS ROM IS = 0FC6'
no ERRORs, 696 Labels, 75D7h bytes not used. Program LWA = EFEFh.
0FC6 is the shipped listing’s own number, and every record the build emits
is byte-identical to the shipped hex — 0 differing records, checked record by
record. That is 3759 of the shipped image’s 4096 bytes verified identical,
and the 337 bytes that are not are the 16 records §7 loses: unverified, not
verified. So ACT 3.0 plus actdbe.py reproduces a shipped Osborne ROM as far
as its output survives, and the DBE/DCE recovery is sound.
Why that is checkable for rev B and not for 1.3 is the whole point. The
rev-B disk carries the answer: OCCROM0B.COM, 4096 bytes, CRC32 f5f697c7.
So the damage shows up as a discrepancy against a reference, and what it ate can
be read off that reference. Our own build of the same source is CRC32
5f48a26a, and the 337 bytes between the two are exactly §7’s damage — nothing
else differs. For 1.3 there is no reference. The same defect eats the same
kind of bytes and nothing can tell us what they were.
So the toolchain is verified for rev B and only trusted for 1.3, and that difference is not a property of the toolchain — it is a property of whether the answer already exists. A control build is a check, not a generator: it can tell you that a build is right, and it cannot supply the bytes a build failed to produce.
There is a second, independent blocker: the 1.3 build carries three errors that the rev-B build of the same code does not. Even with a clean output path the 1.3 image would not be right until those are found.
6.1 The boot test, which the issue makes the pass/fail
It runs, and it says no. Two things had to be learned to run it at all:
o1keys.shsegfaults whenever-biosis passed — with any ROM, including the stockrev1.40.ud11. MAME itself runs fine with-bios(exit 0), so it is the harness’s VRAM reader, not the guest. The way round it is the onedocs/mame-emulation.md§4 already uses: put the candidate ROM in as the default BIOS file (3a10082-00rev-e.ud11) in a scratch rompath and pass no-biosat all. MAME printsWRONG CHECKSUMSand continues.- A screenshot is the only readable output, since the harness cannot be
used. A four-line Lua script —
emu.register_frame_done, thenscreen:snapshot()— writes a PNG that can be looked at.
With that, booting a scratch rompath whose default BIOS is the vendor-built
rev-B image (OCCROM0B.COM, CRC32 f5f697c7, extracted from OCCROM.IMD)
gives its own sign-on and prompt:
LOSBORNE SYSTEM ONE.M
REV B 227
INSERT DISK IN DRIVE LAM AND PRESS RETURN.
That is a boot, and the banner matches the rev-B source’s own ;VER = REV B /
;DATE = 227. (The stray Ls are the ROM’s
DB ESC,'L' / DB ESC,'M' video escapes printing literally; rev B escapes with
ESC+letter where 1.3 and 1.44 use single control bytes, and MAME’s osborne1
video model does not decode the older form. Unexplained beyond that, and it is
recorded rather than guessed at.)
Booting our built 1.3 image the same way gives a full screen of block glyphs and no banner. So the 1.3 image fails its own pass/fail — the check #205 asks for, and it fails it — and the failure is consistent with the two defects above rather than with anything about the source.
What stops the comparison passing outright is §7: the emulated write-back
damages six hex records per build, and the damage is identical on every run, so
neither --recover-ff nor a union of runs can put them back. That is now the
only thing between us and a certified rev-B reproduction — and therefore
between us and a certified 1.3 image.
The 1.3 build carries three further errors that are not yet identified. It
is not the translation (the rev-B build has none), it is not .siz (ACT 3.0
handles the occtxta.ast that has it, and occtxt6.ast does not), and it is
not ~ (NOT) or the RRC/DAA/ADC A,40H instructions — all three were
tested directly. Its LENGTH comes out 0FFA, inside the 4 KB window, so no
overflow error is firing.
So: a 1.3 image can be built, and rom13-reassemble.sh builds one, but it is
not the shipped 1.3 ROM and must not be presented as one.
ACT 3.5E is the clean answer and it is unobtainable — the maintainer has
said so, and it is recorded here so nobody goes looking again. It is in no
local archive (nine ACT80.COM copies checked, plus the schorn.ch act.dsk)
and not on bitsavers. That matters because it means the two defects below have
no clean fix available, which is the whole basis of the recommendation in
§10.
What is left, in order of how much it would buy:
- The output damage in §7 — locate it and the rev-B control build passes, and the 1.3 image falls out of the same toolchain.
- The three 1.3 errors — a source-level difference from rev B, still unidentified; the rev-B build has none.
- The
OCCTXTrevision contemporary with 12/16/81, in case the equates differ fromocctxt6.ast(the choice is inert for this source, but a later revision is not guaranteed to be). - The
OCCRAM2.ASMfrom the 1.3 era — the copy used here is from 10/9/81.
7. A tooling hazard: MAME’s floppy write-back
The assembler’s hex output arrives on the floppy image with runs of
0xDE 0xC0 where hex digits should be — about six records per build, each
one truncated mid-record with the next record’s text running on behind it:
:1001C00\xde\xc0\xde\xc0\xde\xc0…:1001D00…
acthex.py reports them as not a usable record; a clean build has none.
Five measurements narrow it down, and the last two are what close the
question:
- the source files and
ACT80.COMread back off the same floppy byte identical to what was written, so the disk is not generally corrupt, and the guest’s read path is sound; - three runs of the same build are byte-identical files (
md5equal) — deterministic, not flaky, so a union of runs recovers nothing; - the damaged bytes land at the same output-file offsets in both builds (0x480 in each) while the object addresses underneath them differ (0x1C0 for rev B, 0x1F0 for 1.3) — so it follows the output stream, not the object;
- every damaged run starts on a 128-byte boundary inside its sector, i.e. a CP/M record boundary;
- it is identical on a single-density disk (256-byte sectors) and on a double-density one (1024-byte sectors), and on drive A or drive B — so it is not the disk format, the sector size, or the image layout.
That leaves the output path above the disk layer: the bytes are already
wrong when they are handed to the emulated controller. It is the same class as
#203’s one damaged record — acthex.py’s --recover-ff exists for the shape
seen there — but this is six records, and their content is not FF, so that
recovery does not apply.
7.1 What the ROM 2.0 build measured about the same damage
The ROM 2.0 build (#247) hit this again, in the C0 DE shape acthex.py names,
and those measurements narrow it further: they rule out the disk, the drive and
the host’s speed, and leave the build’s own byte stream:
- the damaged bytes are in the raw image the emulator wrote back, not in the
read path:
imdinfo.pyfills an unavailable sector withE5, and these runs areC0 DE, so the image already held them (measured onrun_b.raw); - sending the hex to the other drive changes nothing. The same record was
damaged with
ACT80 B:ROM2 H=B:and withACT80 B:ROM2 H=A:— a different disk, a different allocation, the sameC0 DEat the same record; - throttling changes nothing. Five
-nothrottleruns and one run at the machine’s real speed (O1_MAME_EXTRA=) produced byte-identical damage — the same result §7’s five measurements give for the 1.3 build; - what did change it was the source. The preceding build of the same code — with
two bytes of inline prose in
rom2.diffthat emit nothing — had no damaged records at all, and moving that prose out of the diff’s code lines (so the assembledROM2.ASMwent from 90676 bytes to 90607, both inside the same 177 CP/M blocks) brought the clean build back. The damaged record is the same one either way, with the same bytes under it.
So the variable this build can control is the size and content of what the guest
writes, not where it writes it. What that means in practice is recorded where it
bites: rom2asm.sh’s step 8 refuses the build rather than padding the hole, and
rom2blob.py refuses it a second time — so the next person to hit this gets a
failure, not a silent artifact. acthex.py --merge and
rom13-reassemble.sh’s O1_RUNS remain the union-of-runs answer, and they only
help once the damage moves between runs, which for this shape it did not.
8. What upstream gets, and what it should be told
MAME wants osb13.bin, 4096 bytes, inside osborne1.zip. We can hand over
the source, the method, and the two things we cannot supply. The honest
note to the MAME devs is:
Osborne 1 ROM 1.3 is not a dump — we do not have a 1.3 machine. It is reconstructed from Osborne’s own source, which survives on the OCC engineering source disk
OSROM13.IMD(bitsavers; also in the o1next repo), read as an 8-inch 77-track FM/MFM disk with CP/M skew 3. The source is 10 modules (OCCROM0..9.ASM) linked byOCCROM03.ASM, plusOCCRAM2.ASMand theOCCTXTsystext fromOSUTLSRC.IMD. The source needs SORCIM ACT 3.5E — it uses theDBE/DCEpseudo-ops, which 3.0 and 3.2 do not have, and the shipped rev-B ROM onOCCROM.IMDwas itself built withver 3.5E. We recovered what those two pseudo-ops emit (each byte is the value shifted left one;DCEsets bit 0 of the last byte) from that disk’s ownOCCROM0B.COM, and can build a 1.3 image with ACT 3.0 and a mechanical source rewrite. The toolchain checks out: a control build of the rev-B ROM from that same disk gives no errors and reproduces 3759 of the shipped image’s 4096 bytes byte for byte — the other 337 are lost to a defect in the emulation we build under, which damages a handful of hex records on the way back out. So the 1.3 image is not certified, and it should not go into a romset as if it were a dump.So: this is a reconstruction recipe, not a dump. If you have ACT 3.5E the rewrite is unnecessary and it finishes in one run — we have stopped looking for it ourselves, and we will update this note if it ever turns up.
The point of handing that over rather than a hopeful 4 KB blob is that a wrong
osb13.bin in a romset is worse than a placeholder: it would be
indistinguishable from a dump afterwards.
9. Reproducing this
# the source, read out of the disk (needs cpmtools)
python3 tools/research/o1imd8.py research/source-code/OSROM13.IMD /tmp/rom13.raw
cd /tmp && cpmls -T raw -f occ8 rom13.raw # 19 files
# the whole build, plus the rev-B control, under MAME
tools/research/rom13-reassemble.sh --validate /tmp/o1-rom13
# what a build reports
grep -E "ERROR|LWA|not used" /tmp/o1-rom13/act.log
rom13-reassemble.sh writes /tmp/o1-rom13/OCCROM03.HEX (ACT’s own output),
os1-13.rom (the 4096-byte image), and — with --validate — revb.bin for
comparison against the shipped OCCROM0B.COM. It exits non-zero when the
control build does not match, which is the honest state today.
10. What this closes, and what it does not
| Status | |
|---|---|
research/source-code/*.IMD are readable |
Closed — tools/research/o1imd8.py; all seven disks, not just 1.3 |
| The 1.3 source is extracted and indexed | Closed — research/source-code/ROM-13-source/ |
DBE/DCE semantics |
Closed — recovered and verified against the shipped rev-B ROM |
MAME’s osb13.bin |
Open, and recommended to stay so — see below |
The other NO_DUMP slots (A, 1.2, 1.2.1) |
Open, and closer than they were — OSBROM.IMD, OCCROM.IMD and OSBIOSSR.IMD hold the rev A/rev B source sets and OCCROM.IMD even carries a shipped rev-B image (OCCROM0B.COM, 4096 B). Whether that image is one of MAME’s missing slots is a separate question and is not answered here. |
Recommendation: stop here
We do not have a 1.3 ROM, and the side quest should stop at this point — “record it and stop”, which is what #205 allows for.
The reason is not that the source is short of anything. It is that the two things still in the way are tooling defects, not knowledge, and the clean fix for both is an assembler that is unobtainable:
- the hex output is damaged before it reaches the disk (§7), and
- three errors in the 1.3 build are unidentified while the rev-B build of the same code has none (§6).
Against that, the durable half is already banked: the disk reads, the source is
committed, the DBE/DCE semantics are recovered and verified, the control
build reproduces 3759 of a shipped Osborne ROM’s 4096 bytes exactly, and the
recipe is written down. None of that needs a 1.3 ROM to be useful, and no project path does
either — we build from the 1.44 source.
What would justify picking it up again: a clean output path (a fixed
emulation, or any environment where ACT’s hex file survives), or the
OCCTXT/OCCRAM2 of the right vintage. Both are cheap to test and both are
written down here. Until then, more hours buy nothing, and a wrong osb13.bin
in a romset would be worse than the placeholder it replaced.
Related: #72 (the ROM source index), #70 (ROM inventory), #203 (the 1.44
reassembly and the ACT80 safety copy), docs/o1-rom-variants.md.