on_replace: set other_table and set_fields in flecs_invoke_replace_hook

[Indra-db]

Fix: Populate other_table and set_fields in on_replace hook iterator

Problem

on_replace callbacks had no way to tell whether prev held a valid existing value or uninitialized memory. it.other_table was always NULL and it.set_fields was always 0.

Several call sites also passed r->table after flecs_ensure had already moved the entity, so the table reflected post-move state — making new adds and replacements indistinguishable.

Fix

flecs_invoke_replace_hook — new prev_table parameter

it.table serves two purposes: component storage access (used internally by flecs_on_replace_parent to find name data) and the informational other_table field for user hooks. For the entity-init path these must be different values, so they are now decoupled via an explicit prev_table parameter:

it.other_table = prev_table;
it.set_fields  = (prev_table && ecs_table_has_id(world, prev_table, id)) ? 3 : 2;

it.table stays as the final committed table (required for internal storage access). it.other_table gets the pre-move table. set_fields = 3 means both prev and next are valid (replacement). set_fields = 2 means only next is valid (new add).

Pre-move table at all call sites

• flecs_set_id_move, ecs_set_id, ecs_cpp_set, ecs_cpp_assign: capture prev_table = r->table before flecs_ensure / flecs_get_mut, pass as prev_table.
• flecs_cmd_batch_for_entity second-pass: pass start_table (table before any commands in the batch ran).
• Deferred existing (EcsCmdAddModified): passes r->table — entity already has the component at this point, so set_fields = 3 is correct.
• Entity-init (ecs_entity() with .set): captures init_table = r->table before any components are added, passes as prev_table.

flecs_copy_id refactor

flecs_copy_id previously contained its own hook call using r->table (always post-ensure). Removed it — flecs_copy_id is now a pure copy-and-notify function. Both call sites fire the hook before flecs_copy_id with the correct pre-move table. The non_trivial_set fast path is safe: it is only false when no non-trivial hooks including on_replace are registered.

Semantics

Path	set_fields bit 0
ecs_set / ecs_cpp_set, component not yet on entity	0 (new add)
ecs_set / ecs_cpp_set, component already on entity	1 (replace)
ecs_cpp_assign (requires component to exist)	1 (replace)
Deferred existing — EcsCmdAddModified (fires at insertion)	1 (replace)
Deferred batch flush, new component	0 (new add)
Deferred batch flush, existing component	1 (replace)
Entity-init with .set values	0 (new add)

Tests

All tests register on_replace with a raw hook that captures it->other_table and it->set_fields, then assert both match expected values.

Core (ComponentLifecycle): on_replace_other_table_new_entity, on_replace_other_table_existing_entity, on_replace_other_table_set_existing, on_replace_other_table_batched_new_entity, on_replace_other_table_batched_existing, on_replace_other_table_entity_init.

C++ (Entity): on_replace_other_table_set_new, on_replace_other_table_set_existing, on_replace_other_table_assign_new, on_replace_other_table_assign_existing.

Compatibility

No public API or ABI change. flecs_invoke_replace_hook is an internal function. Callbacks that ignore other_table and set_fields are entirely unaffected.

[Indra-db]

had to add the param for a set fix, if this is not acceptable, you can close this. I don't see any other way around it.

[SanderMertens]

Invoking ecs_table_has_id is not feasible here as it would add way too much overhead to on_replace hooks.