[aliromman]

This constructive interface result replaces the saved-frame oracle identified by PR #5329. A delimited control tape locates and erases its own frames, restores all sixteen payload heads and role-presence bits, and returns arbitrary unequal descriptors. The stack reads no saved parent pointer, frame length or absolute origin. Its eighteen-tape, seven-symbol construction works for every finite properly nested call tree under the stated retirement ABI. The complete kappa calculation is unchanged; selected kappa remains 60403/20000000.

Proof and exact invoice, sequential implementation, receipt, and source pins are fixed at reviewed head ef1cc26e1e9a08c11cf382de948ee433bebab16d. Base and consumed source pins are 7c472563474e60185e6ca101a17622133d047e56.

With k payload roles, parent extents summing to S, parent head offsets summing to H, descriptor lengths d and q, and A absent roles, the literal push/pop total is 13H + 4S + 31k + 18d + 5q + 2A + 39 reads, writes and unit moves. Since H<=S and A<=k, this is linear in current volume and descriptor sizes, independent of ancestor length or depth. Unary offsets are produced by actual head movement; zero has an explicit terminator. The control frame occupies H + 2k + d + 3 symbols. Actual descriptor-size and volume bounds remain premises when absorbing this into a specific node invoice.

The source adapter and limits execute the unchanged retained compact and monomial primitives. Its weak-reference registry simulates lexical allocation and is never consulted by push/pop; allocation is still a separate physical obligation. Empty resumed-role release has its own regression. This PR does not implement recursive C, designated child-output delivery, the general descriptor producer, a complete physical allocation controller, or full supplier admission.

From the repository root, with Python 3.12 and the standard library:

python3 -I -W error -B research/delimited-continuation-stack/verify.py
python3 -I -O -W error -B research/delimited-continuation-stack/verify.py
python3 -B scripts/research_package.py --package research/delimited-continuation-stack

Expected marker: PASS_DELIMITED_CONTINUATION_STACK. Registered normal/optimized CPython 3.12.13 and optimized 3.14.7 checks pass. Evidence covers 1,452 exact small invoices, nesting to depth 31, six complete two-level compact-source cases with all sixteen parent roles occupied, 1,024 actual-recipe phase records and inverse checks, six malformed-state controls, and 32 empty-role lifetime cases. Ancestor lengths 16,257,4096 give identical operation counts in both source layouts. The checker discards the push object and constructs a new stack object before return. Committed-byte checks, the official 17-file export and Git-free optimized replay pass; source and receipt corruption reject. Optional full-history verification was unavailable in the shallow checkout. No full production rectangle, selected cold replay, native compiler, Lean or Docker run is claimed. Hosted CI and independent human or proof-kernel review remain separate.

The advance claim takes the released control-stack handoff. Final overlap covered 305 open PRs, 4,634 Discussion records and all 7,267 tracked README/TASK notes. A concurrent claimant released the same scope after finding this earlier claim. Canonical main still matched the base; arbitrary diffs and unpublished work remain outside overlap coverage. No successor scope is reserved by this PR.

Credit @JoyboyBrian for the segment refinement and the PR #5041 locality invoice; @jmbarrancoml (Jose Manuel) for PR #2250's sequential primitives; Douglas Colkitt for guard/repair; Cinna Davis for the caller; @Matarener for the cover/stock lineage. Their roles and retained notices are documented in the pinned predecessor. New stack proof, code and checks: @aliromman with substantial OpenAI Codex assistance. Apache-2.0; author-checked, awaiting independent review.

[jwr404]

Independent review (James Reed, with Claude Code). Head ef1cc26e1e9a08c11cf382de948ee433bebab16d.

Verdict: correct. Restoration is exact at every nesting depth, and the README's exact push/pop formulas hold, under an independent randomized nested harness.

Harness

I imported only Tape, Counts, Stack and emit from this PR's stack.py; the driver and expectations are mine. Over 300 top-level scenarios:

• States. 16 roles with random presence. Absent roles are empty, as the contract requires. Ancestor prefixes are arbitrary, deliberately including stray LEFT symbols. Parent words have length 0–6 with random heads, descriptors length 0–6, and the control tape starts with an older END or a bare BASE.
• Child work. Each child writes words on present roles and a child descriptor (length q). With probability 0.6 it makes a properly nested inner call, up to depth 4. Then, as the ABI requires, the caller retires the child payloads to their origins before pop.
• Checks. At every frame and depth, all 16 payload tapes plus control and descriptor (cells and head) must be restored exactly. Every frame's measured costs must equal push = 4H + 2S + 10k + 7d + A + 11 and pop = 9H + 2S + 21k + 11d + 5q + A + 28 exactly. H and S are measured from each frame's own LEFT origin.

Result: 1,360 checks (680 frames), 0 failures. My first harness draft measured nested H and S from the tape's original floor and produced spurious mismatches; correcting that to the frame's own delimiter, which is your definition, made every count exact.

The registered replay passes locally (outcome=partial, as declared).

Comparison with #5364 (same handoff)

Both PRs solve PR5329's saved-frame-oracle question with a delimited control tape and unary head headers. For maintainers choosing between them:

	#5356 (this)	#5364
tapes / symbols	18 / 7	19 / 9
header	unary offset from left (h_i)	unary distance to end (s_i − h_i)
child payload cleanup	caller must retire before pop (excluded from bill)	pop erases child bodies (included as 8W)
child descriptor	cleared by pop (5q + 4)	cleared by pop
exact total	13H + 4S + 31k + 18d + 5q + 2A + 39	22d + 10H + 18k + 29 + Σ(2g_i + 5t_i + 7) + 2g_D + 5e + 7

Neither dominates. The bills differ in what they charge: here child retirement is a caller obligation, in #5364 it is included. A cross-reference in each README would help whichever one becomes canonical.