Add a QQ16 ordinary-source tap witness repair

research/qq16-ordinary-source-tap-seed3794b3/NOTICE.md

@@ -0,0 +1,6 @@
+Apache-2.0. This contribution retains predecessor attribution and pins exact source blobs in SOURCES.json.
+
+- `references/pr2093/verify.py` is copied unchanged from PR #2093 at commit `8b21a900eb520e4f11ea177b310f66316f8e8983`; its source header attributes the QQ16 hidden-witness replay and substantial Codex assistance.
+- `references/pr2216/verify.py` is copied unchanged from PR #2216 at commit `b80e9ba43367fb8b20877a97da1cb677c34f9264`; its source header and accompanying `NOTICE.md` retain its original attribution and license.
+- `verify_b1_tap.py` is an adaptation of the PR #2216 verifier with the stated U2-to-T21 tap and read-only receipt behavior.
+- The QQ16 source maps, witness data and retained-tail inputs remain pinned to the scientific base listed in TASK.json and SOURCES.json.

research/qq16-ordinary-source-tap-seed3794b3/README.md

@@ -0,0 +1,105 @@
+# One ordinary-source tap repairs the QQ16 h witness
+
+**New local interface mechanism; no full controller or κ claim.** On the exact
+three-port ordinary-source witness from [PR #2093](https://github.com/CrocSwap/integer-mult-bounds/pull/2093), keep the sum-preserving
+`X_T += X_A; Y_T -= X_A` shear from [PR #2216](https://github.com/CrocSwap/integer-mult-bounds/pull/2216) and add one paired tap at the shared `completed(0)` frame:
+
+```text
+X_21 <- X_21 + X_2
+Y_21 <- Y_21 - X_2
+```
+
+This uses one already-live ordinary register (`U` port 2) and one existing
+selected pair (`T` port 21). It adds no query, helper, recursive child or frame
+transition. The paired operations preserve `X_21 + Y_21` and leave `X_2`
+unchanged.
+
+## Exact result
+
+The pinned hidden witness is
+
+```text
+h_p(d) = c_p bit_0(d b_(p,0))
+c_2 = -1, c_8 = 1, c_85 = 1; all other c_p = 0.
+```
+
+After F0 it is supported on ordinary ports 2, 8 and 85; all selected A/T
+fields are zero. The existing A-to-T shear by itself still produces no selected
+output through the retained tail. With the extra `U2 -> T21` tap, the exact
+full-tail result contains 34 selected outputs, four forward-history values and
+65 completed-X values. In final bank 0, component X, port 21 is exactly
+`11 f_37`; at the identity owner its value is 11. This is a concrete selected
+endpoint that distinguishes the witness from the zero input.
+
+The tap also preserves PR #2216's eight-direction B1 result: the pinned
+post-B1 T-output family remains rank eight, with exact minor determinant
+`-1/4096`. This check ends at B1; those eight directions have not been replayed
+through the remaining tail here.
+
+The search uses the witness support to reduce the one-port/one-T-pair domain to
+3 × 8 = 24 nontrivial taps. All 24 were visible at B1. The first tested tap,
+`U2 -> T21`, survived the complete retained tail, so the search stopped there.
+The other 23 full tails were not run.
+
+## Cost and scope
+
+The added tap costs exactly **2 scalar additions per owner**. Its coefficients
+are `+1` and `-1`, so it adds no coefficient precision. By the triangle bound,
+charge up to a factor of two in the two changed field envelopes (one extra
+magnitude bit) at this local interface. The source is read in the common
+`completed(0)` frame and is not changed. No dirty helper is introduced.
+
+This establishes one exact kernel-witness repair, not full injectivity. Global
+precision, controller coverage, owner placement, restoration outside this
+local state change, all-size admission and the complete recursive moment remain
+open. The selected record and κ are unchanged.
+
+## Reproduce
+
+From the repository root:
+
+```sh
+python3 -I -B research/qq16-ordinary-source-tap-seed3794b3/verify.py
+python3 scripts/research_package.py --package research/qq16-ordinary-source-tap-seed3794b3 --check-only
+```
+
+The first command authenticates the selected archive and both pinned
+predecessor verifiers, checks the no-tap and wrong-source controls, replays the
+whole retained tail, compares the complete exact output receipt, and runs the
+eight-direction B1 control. The package runner is registered for Python 3.12;
+local direct checks use the available Python runtime and are reported
+separately from hosted CI.
+
+## Provenance and coordination
+
+Scientific base: `0fdd7d82de03e43c9e5a1000dfbcd380d1b687d2`; selected-record
+SHA-256: `10b9e340fd8e354a37a0c4813c90e5de1dd42044b5f6aca4a3fc5a42b685fa32`.
+The exact predecessor heads and copied verifier hashes are in
+[SOURCES.json](SOURCES.json). The selected QQ16 maps and witness are pinned
+from the base commit; the two unmerged predecessor verifiers are retained under
+`references/` for reproducible comparison.
+
+Cycle 30 seed: `3794b3df6e4fb4a017103a787d6f6b32`, track
+`recursive-interfaces`, priority revision 4, task `joint-query-controller`
+(refined). The live overlap refresh at 2026-10-15 02:18 UTC covered 482 open
+PRs and 960 Discussion records with complete pagination and no warnings; the
+exact one-port ordinary-source-tap terms had no direct match. The claim and
+handoff are in [Discussion #2205](https://github.com/CrocSwap/integer-mult-bounds/discussions/2205#discussioncomment-18938845).
+
+## Open next questions
+
+- Does the one-bit local envelope fit the complete QQ16 precision and all-size
+  contracts?
+- Do the eight PR #2216 directions remain visible through the retained tail
+  after the new tap?
+- Can the full ordered kernel be made injective with this and other priced
+  existing-field changes?
+
+## Credit
+
+PR #1969 supplies the retained QQ16 maps and local words. PR #2093 supplies the
+three-port hidden h witness and its exact tail. PR #2216 supplies the existing
+A-to-T shear and eight-direction B1 certificate. PR #2079 supplies the earlier
+ordinary-source observation context. New mechanism, integration and replay:
+Cinna Davis with substantial OpenAI Codex assistance. No independent human
+review is claimed. Apache-2.0.