Revision continuation (`RRC-revision-v1`)¶

RRC-revision-v1 is a solve policy, offered by explicit id like T08-warm-v1. It is T06-revision-v2 with one change: when an equation-oriented revision solve ends in a state that edge 3 of the recovery order handles, a third action is tried between the existing two. It is never the default, and no registered policy carries it, so no existing solve changes.

What it does¶

A revision solve that fails on a large edit may succeed if the edit is applied in steps. Under this policy, every revision_eo solve looks up the nearest ancestor revision that has a VERIFIED solve when the job starts. That costs one binding of the ancestor and one diff even when edge 3 never fires. Only its use waits for a failed first attempt: the solve then moves the changed operating values from the ancestor's values to the target's values in fractions of the way (the λ levels), starting from the ancestor's certified root. The first level is the ancestor's own certified problem, bit for bit.

The action order for one failed step is:

continuation in a promoted specification (ADR 0010, unchanged);
revision continuation (this policy; ADR 0072);
the sequential restart (ADR 0015, unchanged).

There is one recovery solve per step, shared by all three. A step does not retry.

When it applies¶

It applies only if every difference between the ancestor's revision document and the target's is a value change of a registered operating specification:

the value, unit or meaning of a pinned state, outlet or duty specification (state.n, state.T, state.P, outlet.T, outlet.P, duty.Q);
the value, unit or meaning of a registered operating parameter, for example a splitter's split_fraction, pressure_drop, efficiency, or a column's condenser_reflux_ratio and reboiler_boilup_ratio — whether it is written as an instance parameter or pinned by a fixed specification whose target is parameters.<name>.

Anything else has no continuation: a connection edit, an added or removed item, a changed model, role or target, a discrete or model-constant parameter (stage_count, feed_stage, nu, damkohler, T_ref, T_scale, coolant_cp), and a first solve with no verified ancestor. The continuation never changes the active set and follows no folds (limitations L20 and L24). The edge is available on the revision_eo route only. On the coupled and legacy routes the policy records absent(not_supplied) and runs the restart. The application looks for an ancestor only under this policy, on a run and never on a rerun.

How a refusal is recorded¶

If any precondition fails (no verified ancestor, a difference that is not continuable, an identity or opening check, a rejected opening), the reason is recorded and the solve falls through to the sequential restart unchanged. The record is:

one initializer_rejected trace event with the message revision_continuation_<status>: <reason>, before the restart's events;
the revision_continuation member of solve-path.json, with status one of absent, rejected, not_reached (the ancestor was assembled but edge 3 never consulted it) or taken, and the reason. It is written whenever the policy is used on a revision solve, whether or not edge 3 fired.

When the action is taken, the trace carries the ordinary homotopy_step events, the certificate's first recovery provenance item records the continuation (type: revision_continuation, the continued parameters, the λ levels reached and the rejected trials), and verification is unchanged: a converged result is VERIFIED by the same independent verifier.

Replay¶

A replay bundle replays alone. replay --rerun takes the recorded revision_continuation member: it does no store lookup and no binding of the ancestor. It rebuilds the candidate from the recorded source parameters, identity, signature and state, repeats the integrity and continuability checks, and runs the solve. The rerun's member must equal the recorded one under its R0 projection; otherwise that is a replay difference. A bundle without the member replays as before.

Measured cases¶

Both cases are regression references (self-generated, with independent checks named), taken on Linux x86-64 and recorded in benchmarks/rrc/results-ref-x86-64.md. In each, the control is the same policy with edge 3 switched off.

Case	Edit	Control	`RRC-revision-v1`	Independent check
N1 (NET-02)	recycle splitter `split_fraction`, source 0.5, target NET-02's registered value	`BOUND_BLOCKED` after 1 iteration	`CONVERGED`, `VERIFIED`, edge taken; 12 iterations; λ levels 1/4, 3/4, 1 accepted, 0 rejected	closed-form worst ratio 0.0109 (limit 1); restart root 1.96e-9 scaled (limit 1e-8)
C3 (M1-recycle column)	recycle `split_fraction` 0.2 → 0.9	`BOUND_BLOCKED` after 4 iterations	`CONVERGED`, `VERIFIED`, edge taken; 22 iterations; 0 rejected	A29 closure 2.73e-10 (limit 1); restart root 7.8e-15 scaled

Under T06-revision-v2 the same two targets also converge (5 and 8 iterations) by the existing recoveries; the policy is not claimed to be faster. The measured recovery cost was 0.12 to 0.63 s per case, with at most 0.051 s to compile each level. Of the six other non-forced candidates, three (N2, N3, C4) qualify and three (C1, C2, C5) converge first under every policy, so the edge is never reached.

Declared limitations¶

L21. Edge 3's continuation exists for a region with a promoted specification (specification continuation). Under RRC-revision-v1 it also exists for a revision region whose nearest verified ancestor has the same declaration and differs only in registered operating values. Everywhere else edge 3 is the sequential restart or unsupported(…). Neither continuation changes the active set. RRC-revision-v1 is not a superset of T06-revision-v2: a continuation that stops part-way ends the step and the restart is not tried (one recovery per step), so it can fail where T06-revision-v2 converges, and no registered case is rescued by it that T06-revision-v2 does not rescue.
Not a superset of T06-revision-v2. On every measured case where this policy's continuation converged, T06-revision-v2's sequential restart converged as well, in fewer iterations. A continuation that stops part-way ends the step, and the restart is not tried (one recovery per step). This policy can therefore fail where T06-revision-v2 converges. No registered case is rescued by it that T06-revision-v2 does not rescue.
L-COL-9. In a flowsheet a column region has edge-3 continuation only under RRC-revision-v1, from a verified ancestor that differs only in operating values. A first solve, a structural edit (stage count, feed stage, a changed end-specification choice) and the default policy have none. The unit's own evaluate keeps the standalone path.
The nearest verified ancestor wins; no further ancestor is searched.
A column with a pinned (non-two-phase) source signature has not been measured: no registered case has one.
The evidence is two registered pairs on one architecture; the aarch64 pairs and the numerical review are pending.

Revision continuation (RRC-revision-v1)¶