fix(session): re-derive from scratch on steps without a delta callback

[justinjoy]

Fixes #1994.

Problem

wirelog_easy_step() on a session with no delta callback re-derived every rule on every step, on top of the rows it had already derived:

• Each step appended another copy of every derived row (3 rows became 6, 9, then 12).
• A removal never retracted the rows it had derived.
• A step with nothing pending re-ran every rule, so a value-position @call ran once per derived row on each idle step.
• A snapshot taken afterwards emitted that state. This is the "step then snapshot duplicates rows" pattern that wirelog-easy.h warned about.

Fix (col_session_step_impl)

• An idle step returns early with or without a callback.
• A plain step clears the derived relations and resets every stratum frontier before re-deriving everything, under the same has_evaluated condition the snapshot path uses. Without the frontier reset, a recursive closure lost (1,4). The full mask is forced even when the pending input recorded only some strata.
• The step sets pending_full_input_eval before clearing, so a failed step leaves a full re-evaluation pending.
• A plain step no longer seeds $r$ retraction deltas left by a removal made with a callback installed.

Rule-head input

Clearing derived rows also cleared input: inline facts (reach(1). beside a rule for reach) and host rows inserted into a rule head. The snapshot path already had this defect on main: a second snapshot after an insert returned reach empty, and so did four-worker recursive steps.

A rule head now keeps its input in a private $in$<name> shadow:

• Created on the first insert into the rule head. Inline facts are loaded through that same insert.
• Inserts and removals keep the shadow in step. A removal planned before the relation is touched is applied after the commit; it is applied even when the relation no longer holds the row, and in that case it schedules nothing.
• Restored after both resets: col_session_clear_idb_rows and the multi-worker recursive reset in eval.c.

Placement and cost:

• Creating shadows at session creation would raise the creation memory floor that test_session pins. Creating them at the first evaluation would leave a reservation behind after a refused step.
• A direct-mapped identity cache keeps inserts into plain input relations at main's speed. 1M single-row inserts: 0.51 s vs 0.51 s; alternating two relations: 0.53 s vs 0.52 s. Without the cache it was 0.71–0.74 s.

Tests

• New test_easy_plain_step (15 cases):
  • idle steps;
  • two-strata inserts and removals;
  • a recursive extend and cut;
  • insert and removal transitions when the callback is cleared;
  • inline facts and host rows in rule heads, on 1 and 4 workers, and via snapshots alone;
  • removals with a callback installed, and after it is cleared;
  • a row inserted twice and removed twice.
• test_extension_replay: no-callback addon counts, a failing plain step followed by a snapshot, and an input-only removal that invokes nothing in another stratum.
• Mutation checks: 14 mutants, each fails a named check.

Docs

• wirelog-easy.h documents both step modes and drops the step-then-snapshot warning.
• docs/SEMANTICS.md specifies the no-callback invocation count and updates its state map.
• Example 12's comment and README are corrected. Checked on that program, step then snapshot returns the same five rows.
• CHANGELOG has two Fixed entries.

Validation

• meson test -C build: 433 OK, 0 failed (before rebase).
• ASAN/UBSAN on both focused tests, and TSan on test_easy_plain_step including the 4-worker cases.
• After rebasing onto 84487d59, the abi suite has one failure: python_encoding flags scripts/perf/hosted_perf_calibration.py, which 84487d59 added. This PR does not touch it.

Not changed

These behave the same on main:

• With a delta callback:
  • the first step publishes an inline fact as a retraction and leaves the relation empty;
  • a host row inserted into a rule head is also published as a retraction.
• If a session's very first evaluation fails partway, the next evaluation does not clear, because nothing has been evaluated yet.

Peer review: an independent reviewer did four rounds. Rounds 1–3 raised blocking findings, which are fixed here: failure recovery, staged removals, rule-head input loss, doc wording, and an over-eager re-evaluation on input-only removals. The final tree had no blocking findings.