fix(ai): a session that has asked forty different questions is not idle

[tpouyer]

Closes #416.

GATE-PROGRESS-1 counted writes the workspace accepted and suite runs over a change set not run
before. Everything else counted for nothing, so these two sessions were the same number — zero:

calls: 40   distinct: 40   repeats: []   files touched: 15     <- run 34402407065
calls: 40   distinct:  1   repeats: [the same read x40]

The first is the second /implement on #410. Its forty calls are in the evidence case the run
filed: search_code {"mode": "map"}, then CommandValidate and staged_refusal by name, a
skeleton of command.py, then strategy.py at four offsets, sandbox.py, ruff_adapter.py,
detected.py, test_validate_phase.py in three windows, the GATE-VALIDATE and GATE-SANDBOX
rows, .lockstep/lockstep.py, and cli.py's ls. On turn 38 it said:

"Now I have a thorough understanding. Let me write the failing test. Here's what I need to test:

1. CommandValidate declares EXECUTES_CODE (currently…"

Two confirming reads later the ceiling stopped it, at $3.46, having repeated nothing.

Why not raise the number

The row records this failing once already — run 34152787128, stopped at turn 20 "still reading the
code it was about to change" — and diagnoses it in its own words:

because the number was set from the clean run's 68 turns without knowing how many came before its
first write

The remedy then was to double it. This is the same defect one doubling later, and that sentence
explains both. The row's own objection to raising ceilings ("gives a wandering run more to use and
a good run nothing") is correct and is the reason this changes what is counted instead.

The change

The tool runner keeps a second tally, explored: distinct read_file, list_files, search_text
and search_code calls, keyed on the arguments. Offset 1 and offset 600 of one file are two
questions — that is how a model reads something too long to hold in one call, and every ranged read
in the run above was one of them. Counted on the asking, not the answer: a refused read and an
empty search both told the session something it did not know.

Before the first write, a turn that asked something new resets the idle counter. It
deliberately does not set last_progress, which is what doubles the allowance — setting it there
would halve the allowance in the middle of the phase the doubling is for, and
test_orientation_does_not_halve_the_allowance_it_exists_for pins that.

After the first write, nothing changes. Twenty turns of reading over a change already staged is
wandering however many different files it opens, and that is the half the #319 evidence supports.

This removes an asymmetry nobody chose: staging a stub on turn 3 bought effectively unlimited
orientation, while reading first hit the wall. The successful /implement on #400 made 44 reads
and 66 searches and survived — because it staged early, not because it explored less.

One pre-existing test moved by exactly one turn

test_gate_progress_1_a_reading_phase_gets_twice_the_ceiling_before_the_first_write expected six
calls and now sees seven. That is the change working, not a regression to paper over: the session's
first read of a file is a question it had not asked, so it is not an idle turn; the six repeats
after it are, and still stop it. The comment there says so and points at the tests holding both
halves.

Verification

• make check green from the committed tree: 2879 passed, 6 skipped; ruff format reports 269 files
  unchanged, so what is pushed is what passed.
• The regression case is run 34402407065's own forty calls, kept verbatim from its evidence
  case rather than paraphrased, replayed through the counter.
• Four falsification controls fired and restored from the commit, each asserting its own edit
  landed first:
  • novelty ignored before the first stage → the regression case blocks again;
  • novelty counted after the first stage too → the post-write wandering test stops stopping;
  • exploring sets last_progress → the allowance halves and the orientation test blocks;
  • the question key forgets its arguments → two windows of one file read as one question.

Objectives

O7 — progress is arithmetic, and the arithmetic now measures what the row says it measures.
O13 — blocked is never a failure, which holds only while the controls stop runs that deserve
stopping; one that stops a working session a turn before its first write spends the budget,
produces nothing, and teaches whoever watches it to raise ceilings.

[github-actions]

in-lockstep review — intent

succeeded · $0.0608 · 3 in / 1001 out tokens

	location	finding
⚠️	src/in_lockstep/ai/invoker.py:571	The asked_something_new branch guards on not last_progress, but last_progress is set to a non-empty string only after a productive write turn; the guard is correct as stated. However, explored is a snapshot taken before dispatch and compared to the value after, yet explores (the baseli
⚠️	tests/in_lockstep/test_orientation.py:163	test_gate_progress_1_a_session_repeating_one_question_still_stops expects idle_turns == 40 (the doubled pre-staging allowance), but given the bug above, idle would reset every turn once explored has risen above zero from the first call — the session would never stop, making this test a false
⚠️	src/in_lockstep/ai/builtins.py:793	The key for _asked serialises argument values with repr(v) and keys with str(k). If an argument value is a string, repr adds quotes, so {'path': 'foo'} hashes as ('read_file', (('path', "'foo'"),)). That is consistent within a run, but it means the key format is not the natural identity
ℹ️		the lens was given this change's blast radius (2325 chars)

[github-actions]

in-lockstep review — performance

succeeded · $0.0480 · 3 in / 374 out tokens

	location	finding
⚠️	src/in_lockstep/ai/builtins.py:793	_asked set grows unboundedly for the lifetime of a session with no eviction.

[github-actions]

in-lockstep review — security

succeeded · $0.0518 · 3 in / 380 out tokens

	location	finding
⚠️	src/in_lockstep/ai/builtins.py:793	Tool argument values are repr()d into the deduplication key, not normalised.
ℹ️		the lens was given this change's blast radius (2325 chars)

[github-actions]

in-lockstep review — tests

succeeded · $0.0551 · 3 in / 626 out tokens

	location	finding
⚠️	tests/in_lockstep/test_orientation.py:178	test_orientation_does_not_halve_the_allowance_it_exists_for does not assert the session completed all thirty reads.
⚠️	tests/in_lockstep/test_orientation.py:155	test_gate_progress_1_a_session_repeating_one_question_still_stops asserts idle_turns == 40 but the attribute name on the outcome is not visible in the diff.
ℹ️		the lens was given this change's blast radius (2325 chars)