fix(verify): agent-loop output budgets share the simple_text thinking-era default (4096 → 20000)

[gadievron]

Summary

finding_verifier and agentic_enhancer/agent call binding.adapter.complete(...) directly (they need raw content blocks for the multi-turn tool conversation), so PR #242's raise of simple_text's default 8192→20000 never reached them: both loops kept a private MAX_TOKENS_PER_RESPONSE = 4096 — half the pre-#242 default — on the largest units in the scan. On the run that filed #290, 31 of 182 incomplete verifications self-declared "finish call truncated at max_tokens". A truncated finish is deliberately downgraded to verification-incomplete (FN-safe), so the undersized cap silently converted adjudications into needs_review.

Fix: extract llm.helpers.DEFAULT_MAX_TOKENS (20000, the #242 rationale comment moved with it); simple_text's signature default and both agent-loop constants now reference the same constant, so the two budget paths cannot drift apart again. Zero behavior change for simple_text callers; the verify/enhance per-turn cap rises 4096→20000 (a CAP, not a floor — models still stop at end_turn).

Note: this issue was sequenced behind #285/#286/#287 so a budget change is measurable (status honesty, error-vocabulary, fingerprint keying) — those landed in #376/#377/#378.

Scope notes (deliberate deviations from the issue's suggestions, with reasons):

• suggestion 1's "read from configuration" and suggestion 3's PhaseBinding property are architecture/config-surface changes, not fixes — deferred; the shared constant delivers the anti-drift property they were after.
• suggestion 4's full 16-site census guard is not added: the other 4096 sites (llm_reachability, report generator, context_enhancer, etc.) are single-shot simple_text-style calls with their own cost/shape tradeoffs this issue did not analyze; blanket-justifying them would be guessing. The drift test covers the agent loops the issue indicts.
• verify appends raw tool results with no input cap, while enhance caps both and documents why #291 (the input cap on appended tool results) is the sibling gap, independent, and remains open.

The private-run figures (31/182) quoted from the issue are provenance-labeled there; the code defect and the fix contract are fully checkable and test-covered.

Test plan

3 hermetic tests: both agent-loop constants ARE the shared constant (identity, not coincidental equality); simple_text's signature default is that constant; no private numeric budget pins remain in the agent-loop sources.

Verification evidence (commands + results)

check	command	result
new + helper tests	pytest tests/test_issue290_output_budget.py tests/test_llm_helpers.py tests/test_llm_helpers_unit.py -q	24 passed
mutation smoke	fix reverted → new test file	errors (import DEFAULT_MAX_TOKENS fails) — RED proof; re-applied → green
hermeticity oracle	env -i HOME=/tmp/fakehome-noconfig python -m pytest tests/test_issue290_output_budget.py	3 passed
full suite	pytest tests/ -q	3152 passed, 2 failed (pre-existing zig local-env, stash-verified), 32 skipped
lint	ruff check . (CI scope)	clean

Fixes #290