Release rescoping decision: 5 of the 6 remaining v3.3/v3.4 blockers are not work

[avrabe]

Follow-up to #187, which landed as a PR — so the decision it asked for is buried in a merged PR body and nothing surfaces it. This tracks it.

Measured state

rivet release status, today:

release	accepted	proposed	cuttable
v3.3.0	7	3	✗
v3.4.0	10	3	✗

Six blockers. One is unbuilt engineering work. Five are measurement outcomes or ordering:

blocker	release	why it is proposed
FEAT-064	v3.3.0	AC1 falsified in practice (#123) and repaired by FEAT-077/FEAT-087 — both v3.4.0, both shipped
REQ-020	v3.3.0	the identity half; gated on FEAT-064
FEAT-069	v3.3.0	measured 0.67% proven rate; depends-on FEAT-095 (v3.5.0)
FEAT-057	v3.4.0	convex polyhedra — the only unbuilt work across both releases
FEAT-094	v3.4.0	own residual: "cannot be scheduled until REQ-021 has a path"
REQ-021	v3.4.0	adjudication; first implementation refuted, 6 wrong-verdict paths (#122)

So v3.3.0 is not waiting on effort. It is waiting on claims that the work itself falsified, plus two dependencies that point at later releases.

The three inversions are now gated

FEAT-096 (#194) makes them machine-checked rather than found by hand: a depends-on crossing forward over a release boundary fails CI. The three known ones are allowlisted with reasons, and a stale entry also fails — so they stay visible and must be re-acknowledged rather than fading.

That means this issue has a mechanical consequence: whichever option is chosen, the fix is to delete allowlist entries, and the gate stays red until they are.

The decision

Unchanged from #187, and still yours:

• (a) Move FEAT-064 and REQ-020 to v3.4.0 — recommended. It makes the plan match what the work discovered: the identity property was refuted in v3.3.0's timeframe and repaired in v3.4.0's, so v3.4.0 is where it belongs. Also drops FEAT-069 to v3.5.0 alongside FEAT-095.
• (b) Accept that v3.3.0 ships after v3.4.0. Honest, and version numbers stop being chronological.
• (c) Re-cut v3.3.0 around what is actually accepted, moving the rest out.

I am deliberately not choosing — re-scoping a release is a planning call, not a mechanical one. Everything above is measured; the trade is not.

Refs: FEAT-096

[avrabe]

FEAT-057 measured before building (#199) — this sharpens the decision here.

The table above says FEAT-057 is "the only unbuilt work across both releases". It now has a measured green light, which matters because the rescoping question is partly "is v3.4.0's remaining work tractable?"

On scry's own scry_mcdc.wasm (851 functions, 23,772 program points):

points carrying genuinely-relational octagon facts	4,749 / 23,772 (19.98%)
points where polyhedra could exceed the octagon (≥3 mutually-constrained locals)	2,512 (10.57%) in 83 functions
Diff / Sum constraint split	24,321 / 4,053

It is not blocked on FEAT-095. I expected it to be — 2,165 of the 2,512 wide points are in functions that degrade somewhere, which reads exactly like FEAT-069's 92.5% result. But that is a different claim: gaps are per-pc and write-set havoc widens only the locals a region writes. Measured directly, only 13 of 2,512 have a gap at or before them.

So the two v3.4.0 blockers now look quite different from each other:

• FEAT-057 — real surface, no upstream dependency, buildable now. A multi-slice arc, but ordinary work.
• FEAT-094 — still gated on REQ-021 having a path, by its own residual.

That does not change the recommendation ((a), move FEAT-064 and REQ-020 to v3.4.0), but it does mean v3.4.0's remaining work is work, not another ordering artifact — which is the opposite of v3.3.0's situation.

Caveat carried from the measurement and worth repeating here: 2,512 is opportunity, not realized gain. It bounds the ceiling from above; it does not show any point would in fact improve. The size of the win is unestablished.

Refs: FEAT-057