fix(storage): reconcile inherited append phase bytes before seal #1156

Problem

An ordinary public import can commit and reopen, then a subsequent append fails before sealing with phase read bytes and calls disagree. Existing #951 evidence records nonzero inherited authentication/parent-catalog calls paired with zero seal/shape phase bytes. Source inspection on current main d81acdaf12ca261af012143422672661b48a2283 confirms initialization populates cumulative raw byte/call fields but leaves the corresponding phase-byte fields zero; later seal and shape handling does not make pre-seal progress valid.

This is a bounded repair of #951's existing phase-attribution acceptance criteria. #951 remains the close gate for full deterministic admission, S20, conditional S22 and adjacent-rung qualification. This issue introduces no new benchmark or release gate.

Required outcome

Preserve actual observed parent authentication and catalog byte/call numerators, assign each read to its phase exactly once, and keep pre-seal, seal, shape and resumed-checkpoint attribution consistent. Keep qualification fail-closed; do not reset counters, invent byte estimates, suppress append, relax validation, or scan graph-sized data to repair accounting.

Acceptance criteria

• A direct Rust facade regression reproduces ordinary import → commit → drop/reopen → append with an initial edge, without resetting process counters; validates progress before seal, seals and publishes, then reopens with exact node/edge counts.
• A resumed/interrupted append regression proves checkpoint restoration and newly observed recovery reads reconcile correctly, without omitting or double-counting inherited parent observations.
• Assertions cover raw bytes/calls at relevant boundaries and conserve the actual inherited contributions through subsequent phases. Missing/inconsistent evidence still fails qualification.
• Existing storage-attribution tests and the affected public construction regressions pass. Run the existing equivalent 1x/2x/4x lifecycle and tiny restart tests as the deterministic integration check; preserve a complete failure census before expanding scope.
• No host ladder, S24/S26 qualification, format migration, or broad attribution refactor is part of this repair. Parent test(scale): attribute physical bytes per node and edge before SCALE26 #951 remains open until all of its own outcomes are proven.

Relationship

Native child of and blocker for #951. The known failure is already documented in #951; this child isolates the verified implementation blocker so its focused PR can close accurately while host qualification remains outstanding.