The 11/200 count did not move when KnotTheory#48 merged, falsifying the
assumption that the alexander/conway sign convention caused it. Split by
Delta(1) over the same 200-trial corpus:

    links (Delta(1)=0):  11 mismatch, 87 agree
    knots (Delta(1)=±1):  0 mismatch, 102 agree

Knots are now fully crossing-order invariant; the residual is entirely a
multi-component-link defect. Upstream conway_polynomial reconstructs from
an assumed symmetry about exponent 0 that no link satisfies (even span
gives a half-integer centre), so it returns genuinely different
polynomials for the same link — 1 + z^2 vs 1 — rather than a different
unit. Filed as KnotTheory.jl#51 with the worked examples.

Comment updated to cite that evidence instead of the old vague wording.
The marker stays until #51 lands.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…e instantiate

#83 changed server/Project.toml [sources] from ../../../<name>.jl to
../../<name>.jl but left this workflow symlinking two levels above the
workspace. Every run since has died in 'Instantiate server project' with

    expected package KnotTheory [215268c9] to exist at path
    /home/runner/work/quandledb/KnotTheory.jl

i.e. main has been red for this as well as the conway marker — the path
error hits first, so the axioms suite never even ran.

Symlink now targets $GITHUB_WORKSPACE/../, matching [sources]. Adds a
check that reads server/Project.toml and asserts every [sources] path
resolves to a real directory BEFORE instantiate, so the next drift names
its own cause instead of surfacing as a registry-shaped error.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Closes the Phase A gap on #34. Before this, EXPLAIN was reachable only
over HTTP with raw SQL; the KRL surface the issue actually specifies did
not exist — `explain` was in neither the lexer nor the parser, and there
was no selectivity model at all.

- `explain` is a reserved word and a statement prefix: `explain from … | …`.
- It parses to KRLExplainStmt, deliberately a DISTINCT statement type
  rather than a flag on KRLQueryStmt, so every existing consumer of
  KRLQueryStmt still means "this query will actually execute" and cannot
  be handed a plan where it expected rows.
- ExplainPlan.jl builds the plan shape the acceptance criteria specify:
  an ordered scan + per-stage operations, each filter carrying column,
  comparator, value, `indexed`, and a selectivity estimate. Conjunctions
  are split into separate filter operations, which is what makes the
  DB-7 reorder meaningful later.

Honesty properties, both asserted by tests:

- Selectivity is the stub the criteria permit pending DB-3 (#33), and
  says so: every estimate carries "selectivity_source" => "stub", and
  `plan_is_costed` returns false while any stub remains — so DB-7 cannot
  silently optimise against invented numbers.
- `indexed` is driven by a deliberately short explicit column list. An
  over-claiming list would make it a fake signal, which is worse than
  reporting false.
- A predicate the plan cannot describe (column-to-column, nested
  boolean, call) is reported opaquely rather than given a fabricated
  column.

25 assertions in server/krl/test/explain_test.jl, wired into the
dependency-free CI step alongside the other pure-Julia suites. Existing
lexer/parser/sql/seam suites all still pass.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>