bug: no command can compare a tic past the first refused one

[MarcusKainth]

Where

Driver loop (driver)

Commit

bb31393, and every commit before it.

OS, architecture and ClickHouse version

macOS 15 arm64, ClickHouse 26.8.2.7.

What happened

No command compares a tic past the first refused one, and none can be asked to.
native diff, native demo and native play all stop there, so on a branch
whose first refusal is tic R, every tic after R is invisible to every
instrument the project has. Not hidden behind a flag; there is no flag.

The span is therefore inspected exactly once, by whoever moves the refusal past
it, as a side effect of unrelated work. #517 moved demo3's first refused tic
from 181 to 210 while adding two attack routines, and that is what first
exposed tics 181 to 209 to comparison at all. A real divergence was sitting in
them, at tic 206, already filed as #502.

So the span grows quietly while nobody can look at it, and whether anyone does
depends on which change happens to move the refusal. That is not a review
process.

#534 is the neighbouring defect and is narrower: a refusal suppressed the
comparison of the tics before it, which #539 fixes. This is the tics after
it, which #539 does not address and which no amount of reporting can, because
the rows do not exist: the run stops.

What you expected instead

Something that can compare a span behind a refusal, so that the tics after R
are reachable deliberately rather than by accident.

A run that continued past the refused tic would need its rows treated as
untrustworthy from R on, since the state feeding them is one the statement could
not produce, so the simplest form is probably not "keep going" but "start from
the probe's own row at tic N and run forwards", which is how #502 was
reproduced by hand. That gives a comparable span behind any refusal without
asking the simulation to carry a state it got wrong. Whether that belongs in
native diff as a flag, or as a separate command, is a design question rather
than a preference of mine.

Reproduction

# On any tree whose first refused tic is R, no invocation compares R+1.
clickdoom native diff 2134 --probe <trace>   # stops at R, exit 3
clickdoom native demo demo3 --from sim       # stops at R
clickdoom native play                        # stops at R

Output

$ clickdoom native diff 215 --probe refemu/reference_traces/demo3/probe.9a6a47d01119.tsv
# native diff elapsed=55.4s tics=213 tics/s=3.8
clickdoom: error: tic 210 unresolved: PL_ACTION_NEEDED
exit=3

The run produced rows for 213 tics. Tics 211 to 213 exist in native_state and
are not compared by this or any other command, and nothing reports that they
were skipped.

[MarcusKainth]

Design settled with the maintainer, recorded here so the issue carries it rather than a chat log.

CLI. clickdoom native diff <TICS> --start <TIC>. --start defaults to 1, which is today's behaviour exactly. --from stays a source (probe, sim) on demo and render and is not reused here, so one flag name never means two kinds of thing. --start greater than TICS is refused at parse time, since it would compare nothing.

Seeding. probe::copy_into_state already copies probe rows into native_state; it gains a single-row form and restart calls it instead of writing the level's setup row. A probe with no row at N-1 fails loudly, because seeding from nothing would leave the setup row in place and the run would look like a --start 1 run that agreed.

Bounds. The seeded row at N-1 is the probe's own, so comparing it against the probe agrees by construction. With only an upper bound it would be included, agree, and be counted, so a run that verified nothing would report one tic compared and agreed. parity::first_divergence and field_summary therefore take a lower bound as well as the upper one #539 gives them, and compared() takes the same, so the count reported is the count actually held against the probe. The refusal search starts at --start too.

Report. Says where it started, what it compared, and where it stopped: seeded from the probe at tic 210, compared 211 to 260, every field agrees. A --start 1 run keeps today's wording.

Contract. NATIVE.md §7 gains a sentence saying a seeded run verifies the transitions from the state it was given rather than a cumulative run from tic 1, so the two are not the same evidence. A spec-change issue carries it, drafted with the PR.

Tests. A live arm seeding past today's pin and reporting the next refusal; an arm planting a divergence inside the seeded span; and an arm asserting the seeded row itself is not counted, which is the one that catches the bounds mistake and the one easiest to leave out.

Blocked on #539, which this builds on.