sysinfo|generation: both is a value now, and the report value-sets are declared open

Two points came back from the consumer (orbistoun), both obSCEne's to settle.

both and neither had collapsed to the same unknown. The earlier fix sent both to unconfirmed|unknown and neither to absent|unknown - technically distinct, but unconfirmed means "resolves, no confirmed signature to read it" (IP/FW), which mislabels both (a positive observation) as a flavour of "couldn't read it", and the value was identically unknown. The 005-section already treats the two as distinct partials; the record now matches. both is a value of its own: sysinfo|generation|known|both, distinct from neither's absent|unknown. It still names no console - presence is not implementation - but "both driver families resolve" is a real finding (the stub-everything fingerprint, or real back-compat) and a different fact from an absence. Verified on shadPS4: OBS|sysinfo|generation|known|both, clean run.

"Is the state vocabulary closed?" - decided and written into the contract. It was unspecified: OUTPUT.md's versioning covers new record kinds, appended fields and field meaning/order, but was silent on new values within an enum. Decision: report enum values are open (a reader tolerates an unrecognised value, obSCEne may append one without a version bump - same reasoning as adding a record kind, because the report exists to surface findings); the protocol grammar is the closed exception (verbs, refusal reasons, capabilities are fixed lists the spec and protocol.py enforce). Landed in docs/OUTPUT.md and the handover.

Process change worth keeping

• A shared file replaces relaying by hand. obSCEne and orbistoun were passing feedback by having messages copied between two Claude sessions by hand. That is now <shared>\obscene-orbistoun-bridge.md
  • a neutral log (neither repo owns or commits it) both sides post to. Durable contracts still live in OUTPUT.md / PROTOCOL.md / the handover; the bridge is for the conversation. make check green (117) throughout.

---

Pack/unpack conversions + per-kernel input scheme (D146). The five D124 deferred: packunorm, packsnorm (arity 2), unpackunorm, unpacksnorm, unpackhalf (unary, packed word). They needed range-bound operands, so first built the input scheme: obs_gpu_operands(name) picks the multi-operand set ([0,1] for packunorm, [-1,1] for packsnorm, bit edges otherwise) - the multi-path twin of D124's unary bit-kernel predicate; unpack kernels join the bit-pattern list. One predicate per path, no schema change. 50 kernels; all match the reference bit-for-bit. Census: cvt_pknorm_u16/i16 moved isa -> covered (they were GLSL-reachable all along), and the unpack conversions added to base ISA. unpackHalf HAS a reference (f16->f32 is exact via a bit-exact f16_to_f32 decoder), unlike packHalf which needs f16 rounding. Golden inspected then re-blessed: 133 new lanes, all five new kernels, nothing else moved. 1668 records. verify: ok.

---

gpustats: the diff as a distance (D148). obscene-tool gpustats device reference reports per-kernel ULP, ranked worst first - the question the transcendentals need (how far, not just whether). Analysis tool like gpuref (prints, exits 0, no verify gate; correctness in unit tests). On the llvmpipe golden it named the shape: well-behaved SFU at 1-6 ULP (rsq 1, exp2 2, exp/cosh 6), catastrophic on large-argument trig (sin/tan/cos of ~1e20 land on the wrong side of range reduction). Exact ops show 0.

Surprise (two, both from the first output)

• The naive 0x80000000 - bits ULP ordering put +0 and -0 2^32 apart, making min/max/pow/sinh look catastrophic on signed-zero lanes. They compare equal as numbers, so the mirror-magnitude ordering collapses them to 0 ULP; the sign of a zero is gpudiff's concern, not an error metric's.
• Integer-output kernels (cvti, bitcount, bfe, packs...) read as floats give nonsense distances (a bitcount of 32 as a float is a tiny denormal). Omitted by name, with a comment that a new integer kernel must be added to the list; they are gpudiff's exact-or-bug domain. 123 tool tests, clippy clean; C side untouched so verify not re-run for it.