Improve planning and continue cleaning - #38

explain_physical now prints the fallback root's node kind alongside the
reason (MaterializedCall(Melt: root is not a row-local map)), matching the
physical_plan.hpp contract that the line names the logical subtree the
fallback keeps. Adds plans/physical-fallback-adapter-plan.md for the
remaining Phase 5 item 3 work.

Co-Authored-By: Claude Sonnet 5 <[email protected]>

build_operator_impl's per-kind switch carried 15 branches (Columns, Melt,
Dcast, Cov, Corr, Transpose, materializing Join, Matmul, Update, Resample,
Window, AsTimeframe, Model, Construct/Stream, Program) that each rebuilt a
child through build_operator, materialized it, and called a table
function. Every one of those kinds already has an equivalent branch in
interpret_node, which recurses through itself -- so the switch was a
second copy of that dispatch, kept in sync by hand.

Delete the branches and let the fallback tail (interpret_node +
make_table_source) handle them. plan_physical does not migrate these
shapes, so they are still counted by note_materialized_call, and
physical_fallbacks_for(kind) now buckets the whole backlog by node kind.
A fallback subtree is planned once at its root rather than re-entering
plan_physical per node.

Trade-off: the input side of these breakers no longer runs through the
fused parallel scan-filter physical path (interpret_node copies whole
tables and filters serially). No PDS-H or ClickBench query hits bucket A
-- q13's left join is FusedLeftJoinCount, a migrated path -- so this is
unmeasured on the benchmarked shapes. Recoverable per kind by lifting it
to a migrated breaker-over-pipeline the way Join / Aggregate / Order were;
the per-kind fallback counter is what will drive that.

build_binary_materializing_operator had no other caller and is removed.
All 1,815 non-slow tests pass; strict GCC build and Debug/Release plugin
builds pass. chunked.cpp -356 lines.

Co-Authored-By: Claude Sonnet 5 <[email protected]>

Routing every materializing breaker straight through interpret_node
(previous commit) re-evaluated its inputs whole-table and serial:
join_filter_rank regressed +14.7% because the filter feeding its
grouped-rank update lost the fused parallel scan build_operator gives it.

build_operator_impl's fallback tail now calls build_materialized_fallback,
which builds each of the node's relational inputs via build_operator +
materialize_operator, then runs interpret_node over the node with those
inputs handed back through a new ExecutionContext::pre_materialized_children
list (node ptr -> table). interpret_node checks the list at entry and
returns the pre-built table instead of recursing; only direct inputs are
listed, so the node itself and anything deeper interpret normally.

fallback_relational_inputs is an explicit switch allowlist. Window's
direct child is an `update` clause (interpret_node must own it, with the
window duration) so its grandchild is the input; Stream / Construct /
Program and unlisted kinds get no pre-build (their children are template
or expression nodes). A first cut without those exclusions failed 51
tests by evaluating an update clause or a __stream_input__ transform
standalone.

A/B vs pre-step-2 (c9a20fb1) over join,reshape,window,stats,transform,
multi, 12 interleaved repeats: every query noise, join_filter_rank -3.8%,
total +0.9%. All 1,815 non-slow tests pass; strict GCC and Debug/Release
builds pass.

Co-Authored-By: Claude Sonnet 5 <[email protected]>

Co-Authored-By: Claude Sonnet 5 <[email protected]>

Whole-tree `scripts/clang-format.sh -i` over src/include/tests/tools.
Mechanical only. The drift is almost entirely from the pipeline-extraction
commits (0d60d634 and earlier), which were not format-checked at commit
time; this session's fallback-adapter change added one wrapped line in
interpreter.hpp. format-check CI runs the whole tree, so it gates on all
of it.

Co-Authored-By: Claude Sonnet 5 <[email protected]>

clang-tidy 18/20 (CI) flags std::optional::value() as an unchecked access
when it can't prove the optional is engaged -- clang-tidy 23 (local) does
not, so the previous commit's .value() calls passed here but failed CI.

Add explicit has_value() guards in the aggregate and join operator
methods that read columns_ (they run only after bind_aggregate_columns /
setup_right_emit_schema has bound it; the guard makes that a checked
error rather than an assumed invariant), and check the inner
aggregate_inputs[i] optional before dereferencing it (nullopt only for
Count, already skipped, but clang-tidy can't correlate that). Also
re-applies the pipeline_executor performance-move-const-arg fix dropped
in an earlier revert.

Verified with clang-tidy 20.1.0 via the CI clang-tidy-diff path: no
warnings on the changed lines. All 1,815 non-slow tests pass under Clang
and strict GCC; format clean.

Co-Authored-By: Claude Sonnet 5 <[email protected]>