FIX: warn when selected attack techniques have no registered factory

[fei (feiiiiii5)]

Description

resolve_technique_factories() in pyrit/scenario/core/matrix_atomic_attack_builder.py silently dropped selected techniques whose factory is not registered — the docstring even documented the behavior as "silently dropped". When a scenario (or a user of extra_factories) selects a technique that was never registered, e.g. due to a typo or a custom technique added under a different name, the run quietly proceeds with fewer attacks and there is no signal anywhere that selections were discarded.

This PR emits a single warning naming every missing technique, then proceeds with the remaining ones as before:

WARNING  pyrit.scenario.core.matrix_atomic_attack_builder:matrix_atomic_attack_builder.py:171 Skipping 2 selected attack technique(s) with no registered factory: missing_a, missing_b. Register the technique(s) (or pass them via extra_factories) to include them in the run.

Design notes:

• Warning, not an exception: dropping is intentional fallback behavior (per the original docstring), and callers may legitimately probe which techniques are available; this only makes it observable. Happy to switch to raising if maintainers prefer strictness.
• One aggregated warning per call, names deduplicated but kept in selection order, so repeated selections don't spam the log.

Fixes #2461

Tests

Added three regression tests to TestResolveTechniqueFactories:

• warning is emitted when a selected technique has no factory, and names it
• no warning when all selected techniques resolve
• one warning listing each distinct missing name once, in selection order (deduplicated)

All 35 tests in tests/unit/scenario/core/test_matrix_atomic_attack_builder.py pass locally (73 failed / 569 passed elsewhere in tests/unit/scenario/ is identical to the pre-change baseline on this machine — those failures are environment-only, unrelated modules). Ruff check + format pass on both touched files.

[fei (feiiiiii5)]

Both addressed in a1600c2 (+ merge 0fa4881 pulling in main's jailbreak changes):

1. Authoritative skips: resolve_technique_factories() now returns a TechniqueResolution (resolved factories + skipped names in selection order); jailbreak/adversarial record resolution.skipped and the base Scenario persists it into ScenarioResult.metadata["skipped_techniques"] at result creation. The run summary reads that persisted record — no display-group derivation. Your Rapid Response/Adversarial/Jailbreak examples would previously have mis-reported; a new regression (test_display_group_labels_do_not_affect_reporting) pins that custom labels never change reported skips.

2. Shared fixture: _make_db_scenario_result now builds a real ScenarioResult via model_construct carrying scenario_identifier + metadata (with MagicMock method overrides for get_techniques_used / objective_achieved_rate / get_display_groups since individual tests configure them), fixing the 28 failures. Also merged main's jailbreak fail-fast for missing techniques with this branch's resolver semantics: resolver raises only when nothing resolves; scenario-layer raises on any partial miss per main's newer behavior.

Full local: builder tests 38/38, service suites green, ruff clean.

[fei (feiiiiii5)]

Both gaps closed in 6b67551:

1. build_matrix_atomic_attacks now returns (attacks, skipped); Cyber / Leakage / RapidResponse unpack it and set self._skipped_techniques before the base Scenario persists metadata["skipped_techniques"] — so those scenarios' summaries and CLI output no longer lose partial-miss information.

2. Multilingual consumes resolution.resolved (its dict-style iteration was the source of the 9 TypeErrors) and records resolution.skipped. The Jailbreak mock now returns TechniqueResolution(resolved={}, skipped=[]) per the new contract.

Local: multilingual 14/14 (was 9 failed pre-fix), builder 38/38, service suites green; ruff clean on all touched files.

[Roman Lutz (romanlutz)]

one more minor comment, then it should be ready

[Roman Lutz (romanlutz)]

This function now returns (attacks, skipped) but is still annotated as list[AtomicAttack], which causes four ty errors in this helper and its callers. It also breaks the documented custom-scenario pattern in doc/code/scenarios/0_scenarios.py, where _build_atomic_attacks_async() directly returns this helper and now returns a tuple instead of the required attack list. Please fix the return contract and update both synchronized documentation files, or preserve the existing list-returning API.

[fei (feiiiiii5)]

Addressed in 1b30e56:

• build_matrix_atomic_attacks is now annotated tuple[list[AtomicAttack], list[str]], which also clears the four ty errors (the one in the helper plus one in each of the Cyber / Leakage / RapidResponse overrides that unpack it).
• The documented custom-scenario pattern in doc/code/scenarios/0_scenarios.py and its .ipynb twin now unpacks the tuple and records the skip list on self._skipped_techniques — that is the path that surfaces a partial technique miss in ScenarioResult.metadata and the run summary, so the example teaches the same contract the built-in scenarios use. The scenario contributor instructions (.github/instructions/scenarios.instructions.md) describe the second return value the same way.
• Two call sites ruff-format wants joined, and a duplicated usefixtures marker in the builder tests, dropped.

If you'd rather keep the one-line helper for custom scenarios, I'm happy to preserve a list-returning API instead — but since surfacing the skip list is the point of the change, unpacking felt like the honest shape for the documented pattern.