perf: removing a stage takes stage1 analysis from 40 s to over 600 s

[MarcusKainth]

What kind

Cost grows where the model says it should not

Which benchmark

No make target. clickdoom native demo demo3 --from sim --no-window --times-out <path>, reading the first tic's stage1_us, which is the
statement's own analysis.

The numbers

ClickHouse 26.8.2.7, main at 89e73f9, one throwaway container, machine lock
held. Statement sizes are from the generator, not system.query_log, so they
are not truncated.

  statement                       stage1 bytes   stage1 analysis
  the tic's first statement          2,015,048          39.74 s
  the same, with the last stage
  of mobj::thinkers removed          1,850,587      did not finish

The second statement is 164,461 bytes smaller, 8.2% less text, and does
strictly less work: it is the first statement with the call to
mobj::thrown_thinks left out and nothing else changed. The driver gave up
on it at FIRST_TIC_TIMEOUT, and the server still listed it in
system.processes at 660 s.

  clickdoom: error: a statement failed: the simulation's own first
  statement: the server did not answer within 60s of the body closing
  (system.processes still lists it, 660.0s elapsed)

So removing one stage took the analysis from 39.74 s to over 600 s, a
factor of at least 15, while making the statement smaller.

For scale, the same experiment over eleven other cut points behaves the way
you would expect, monotonically, from 0.10 s for a statement cut after the
tic command to 39.74 s for the whole thing. Only this one cut inverts.

The machine, and how quiet it was

Apple M-series laptop, macOS 25.6. Machine lock held, one ClickHouse
container and nothing else running, load average 2.93 at the start. No other
lane's container was up.

What you think is happening

I do not know, and I am reporting it rather than explaining it.

NATIVE.md already says analysis cost is not a function of the statement's
size and that where it comes from is not settled, so a smaller statement
costing more is not by itself a contradiction of the model. What is worth
recording is the size of the effect: a factor of at least 15 from removing
one stage, where every other cut moved the cost smoothly.

The cut was produced by a local-only switch in bindings_stage1 that
returns the Tic before mobj::thrown_thinks is staged. Nothing else
differs, and the resulting statement passes the same binding-order and
parenthesis-balance checks the committed one does. It is possible the fault
is in that switch rather than in the server, and I could not rule that out
from the outside, which is part of why this is worth someone else looking
at.

This matters beyond curiosity because the projection is 64% of stage1's
analysis (25.5 s of 39.7 s) and is the obvious thing to restructure. If
analysis cost is not monotonic in what a statement contains, then a rewrite
that plainly does less work cannot be assumed to analyse faster, and every
such change needs measuring rather than reasoning about.

[MarcusKainth]

Does not reproduce on 31b0674.

The same cut, the first statement with mobj::thrown_thinks left out, now
analyses in 31.9 s against the whole statement's 40.1 s, across two runs of
the ladder. It is monotonic with every other cut and nothing hangs.

What changed in between is #517 and #520, neither aimed at this. #520 moved
the sector thinkers and put their draw count into stage1 as its own stage,
which changes what sits either side of the cut, so the statement that hung
no longer exists in that form.

I am leaving this open rather than closing it. The original measurement
stands: on 89e73f9 that cut did not finish analysis in 600 s while being
164,461 bytes smaller than the statement that analysed in 39.74 s, and
nothing in #517 or #520 explains why removing a stage would cost a factor
of fifteen. What the disappearance says is that the effect depends on the
exact shape either side of the cut, not that it was never there. If a
projection rewrite lands and analysis moves the wrong way, this is the
first thing to look at.

The cut is reproducible from the tree now: bench::Cut::Thinkers in #542,
timed by the harness in #545.