increment 3: realization materialization arm — reconcile→Share collapse (KeySource-keyed)

[gunbai-bot]

Increment 3 of the realization sequence (the materialization arm), following the increment-1 CreateEffect idempotency vocabulary already on main.

What this does

Re-introduces the trimmed materialization vocab in std.realization and adds the reconcile → Share collapse:

• type Placement = LocalInProcess | LocalFilesystem | RemoteNetwork
• type Materialization = Recompute | Memoize | Share
• type RealizedStep { shape, placement, materialization, cost }
• fn reconcile(steps) — collapses two same-KeySource CreateIfAbsent steps into ONE step with materialization: Share; distinct identities (or non-create effects) stay separate.

The collapse decision is std.effects.create_double_init_collapsible (the single KeySource-keyed authority). reconcile only consumes it — it never re-derives KeySource equality (§3 single authority).

Not Route-C / content-hash gated

The collapse keys on KeySource (create_double_init_collapsible → key_source_eq, pure structural over the closed KeySource coproduct), never on a content hash. The step carries its source EffectShape so the collapse can key on identity now. RealizedStep's intent/licensed_by ContentHash fields return in increment 4 with cross-tree content-hashing (trim marker kept in the header).

Proven by execution

dsl/test/claim/realization_reconcile_witness_test.dag → reconcile_share_collapse_witnesses is green by execution, with a discriminating +/- design:

• GREEN same-KeySource → exactly ONE step, exactly ONE Share.
• RED-if-wrong different-KeySource → TWO steps, ZERO Share (a collapse-on-anything-but-KeySource bug flips this red).
• Fail-closed PostAlways create-always → never collapses (TWO steps, ZERO Share).

Whole-tree dsl compile-clean passes (281 modules, 302 files, 0 diagnostics); existing realization_witnesses (independence arm) still green (no regression).

Latent bug fixed (pre-existing, not new scope)

create_double_init_collapsible called key_source_eq(a:, b:), but the dsl key_source_eq is defined (left:, right:). Latent because nothing in the dsl tree consumed create_double_init_collapsible until now — reconcile is the first consumer, so the whole-tree compile-clean gate would have broken without the fix.

Deferred to later increments

• Increment 4 — generalize realize() + re-introduce RealizeIntent { target: ContentHash } (needs cross-tree Node content-hashing).
• Increment 5 — executor honors Share (the seed change): wire Materialization::Share through the executor / resolved_graph_cache so the CI floor resolves shared closures once.

🤖 Generated with Claude Code

[gunbai-bot]

Addressed in 504bbcc. Took the first option from the review: deleted is_share from std.realization and moved the Share discriminator into the witness as a test-local fn. Rationale — reconcile sets materialization: Share directly and nothing in std/ reads it back, so the predicate had zero std/ consumers; per §6 model-just-in-time it does not belong in std/, and a future increment-5 executor that honors Share will add its own dispositioned detection rather than us pre-staging it now. std.realization keeps only the fold-bearing reconcile arm. Verified by execution: dsl compile-clean (281 modules, 0 diagnostics) and reconcile_share_collapse_witnesses => true on the new head. — sent from merry-owl-830

[gunbai-bot]

Thanks — approval noted. The has_collapsible_peer nit is correct (peers |> length > 0 reads cleaner than !((peers |> length) == 0)). Holding it deliberately rather than pushing now: it is a pure-cosmetic non-blocker, and a one-line touch to realization.dag would re-stale the approval just landed + restart CI while the second review is still queued and cursor’#39;s (now-stale) REQUEST_CHANGES re-review is pending on this head. I’#39;ll fold the cleanup into the increment-4 PR, which revisits realization.dag to generalize realize() anyway — keeping this PR's head stable so the converging approvals stick. — sent from merry-owl-830