D24: recover five EIII/ESI banks whose header name is a mistyped copy of the directory's

[bmxcode]

Closes #43.

What this deliverable did

Five named EIII/ESI banks across three discs claimed a volume, found no bank header for their directory name, and read nothing — each correctly carrying the ADR-0012 note no bank header found for this bank; listed only, but named like ordinary sample banks rather than like the E3 Main Code OS slots that share it. Issue #43 left open which of three each was: a genuinely empty slot, a header the search could not reach, or a second copy of a bank read under a name written twice.

Measured against the discs (SAMPLERDISC_TEST_DISCS), all five are the same thing: a header the search does not reach. Each has a real EMULATOR/EMULATOR THREE header sitting at exactly the address ADR-0021's placement fit predicts, holding real audio named for the bank — but the header's own 16-byte name field at +16 is a corrupted copy of the directory name:

Disc	Directory name	Header +16 name	Records	First record
Vol. 10 – Elements of Sound 1MB	Electric Grand X	Eelectric GrandX	9	ELEC GRAND _000
Vol. 16 – Ditto Drums	PERCUSSION#1 X	PERCUSSION #1 X	31	TAMB BRASS
Vol. 17 – Heavy Guitars	HvyGtr FX5 X	HvyGtr FX5 XX	2	Gtr FX 11
Vol. 17 – Heavy Guitars	Misc Gtr FX 2MbX	Misc Gtr FX 2mbX	6	Gtr Feedback Shr
Vol. 17 – Heavy Guitars	HvGtrFdBkTxtr2Mb	HvGtrFdBkTxtr2M	1	Gtr FeedbackLoop

Every corruption is a shifted space, a case change, or a single doubled/dropped character — an edit of at most one once the name is lowercased and its spaces stripped.

The rule (ADR-0031)

A directory entry that no header names exactly binds the header sitting at the address its placement predicts, when all three hold: a header is there, the header's name is within one normalised edit of the entry's, and no other entry already claims it. _bank_offsets gains this recovery pass after its existing placement arbitration; nothing downstream changes, so a recovered bank flows through the same record walk and the ADR-0012 empty-with-note invariant is preserved.

This is ADR-0021's instrument one step wider — from which of two headers wearing this exact name to the near-named header this entry points at. Nothing is placed: a bank whose predicted address holds no header binds nothing.

Why the gates are load-bearing

The rejected alternative — bind by the predicted address alone — recovers all five but fails on ditto-drums, measured: the real OS-code slots E3 Main Code and E3X Main Code predict addresses that fall on the Ditto Drums X and DAVE W KIT1 X headers (other banks the directory names elsewhere). Address alone would hand each slot another bank's audio under the wrong name. The name gate (e3maincode is a dozen edits from dittodrumsx) and the unclaimed gate reject them together. Also rejected: fuzzy-matching names everywhere (reopens ADR-0015), leaving them noted (withholds 49 real samples), and rewriting the header name (invents an authority the disc does not grant).

What moved

• Vol. 16 – Ditto Drums: 948 → 979 samples (+31, PERCUSSION#1 X), loops and stereo unchanged.
• Vol. 10 – Elements of Sound 1MB and Vol. 17 – Heavy Guitars: pinned in tests/test_discs.py for the first time (102 volumes / 1 465 samples, and 68 / 870).
• No reference disc moves. Simulated across all ten EMU3 reference discs and the three size-twins in the collection, the recovery binds nothing — it fires only where a name was mistyped.

Pins added/updated in _EMU3, _HEAD_DIGEST (the 1MB disc shares its size with the 2MB cut), _EMU3_RUN_ENDS, and a new _EMU3_RECOVERED_BANKS table asserting each of the five by the mistyped header name, not just the count. Synthetic unit tests (ADR-0008) cover _near_name, the recovery/unclaimed/near gates in _bank_offsets directly, and the end-to-end recovery via a new header_names fixture knob.

What it deliberately does not claim

Two of these discs list a bank name twice in the directory — Harpsichord X on Vol. 10, HvyGtr Maj.Open on Vol. 17 — and because located is keyed by name both entries resolve to one header and double-list the same records. That is a different mechanism (it reads the same audio twice, not nothing), one arm of the heavy pair points at a real but blank-named 6.3 MB header, and it needs per-entry header resolution. It is left to #47 and noted in docs/README.md; the D24 pins for those two discs include the double-count until it is fixed.

Verify

uv run ruff check .
uv run ruff format --check .
uv run pytest -q
uv tool install --editable . && samplerdisc --version
SAMPLERDISC_TEST_DISCS=… uv run pytest -q tests/test_discs.py -k emu3