.kiro/specs/pre-push-safety-net/implementation-group-3.md

# Group 3 implementation verification

Date: 2026-09-13. Scope: tasks 3.1–3.4. The maintainer authorized continuing this
group after the group 2 completion and next-step summary. Tasks remain unchecked
until independent review and fresh verification establish their completion.

## Boundary

This group integrates the existing backend and ranker with generic execution,
change-context and scope contracts. It does not implement new detector semantics,
alert admission, persistent caching, or hook installation. Equivalent normalized
PHP and JavaScript evidence must retain equivalent core decisions. Vendor sources
must remain outside both graph inputs and targets; missing external semantics are
coverage limits.

## Verification state

Group 3 is independently approved and VERIFIED (TASK scope): tasks 3.1–3.4 are
complete, bringing the pre-push spec to 11/27 tasks. The approved contributor-scan
frontend-retention prerequisite 2.18 is also verified. Groups 4–8 remain pending;
this is not feature GO, capability admission or a hook latency claim.

## Baseline

Groups 1 and 2 completed at commit `948ee02`; the working tree was clean before
this group began. The prior final suite recorded 1,069 passing and nine skipped
tests. Three pre-existing mypy diagnostics concerned reuse of `candidates` in
`model/scan.py`; fresh checks will report whether they remain.

Historical Joern query failures cannot be declared resolved merely because a
bounded failure returns promptly. Runtime smoke establishes readable source,
nonempty graph census and query payloads for its declared small fixtures. It does
not establish full-workload detection, actionable precision or hook latency.

## Task 3.1 evidence

The backend accepts an optional generic `ExecutionBudget`. Each build binds an
independent session to its graph callbacks; probes, builds, overlays, census,
queries and retries consume the same remaining time. Process startup also counts.
A timed-out stage prevents further retries. The budgeted process path kills the
original process group even after its launcher exits, bounds reaping and owned
scratch cleanup, and exposes incomplete cleanup explicitly. The driver releases
graph ownership on exceptions as well as normal return. Legacy backends cannot
silently discard a supplied deadline; optional unbounded judge calls are withheld
with an explicit reason. Ordinary scans retain their existing budget behavior.

`benchmarks/measurements/2026-09-13-engine-budget-smoke.json` records the locally
rebuilt non-root image and exact packaged backend/driver hashes. Under one
900-second lab budget, native PHP read 45 source bytes and JavaScript read 48;
both produced populated graphs, dataflow overlays and exact source-node witnesses.
Unreadable PHP source was explicitly refused. PHP took 73.620 seconds and JavaScript
34.756 seconds in this small smoke; these are runtime checks, not hook latency.

An independent controlled process probe read its 241-byte input script, launched
a parent and child ignoring SIGTERM, and observed termination in 0.804 seconds
under a 0.8-second deadline plus 0.8-second cancellation allowance. Signal delivery
may briefly leave a child visible immediately after return; the probe polls only
within the original allowance. Zombies count as terminated, not running descendants.
Concrete reusable graph leases and persisted completion records remain later tasks.

The three former mypy errors were resolved by renaming the evidence-region list
inside the driver; this does not change the ranker. Global mypy now checks all
108 source files successfully. Independent review APPROVED task 3.1. Its fresh final suite passed: 1,089 tests,
nine skipped, 96.85 seconds; global Ruff, formatting (219 files), mypy and diff
checks passed. Verification status: VERIFIED, claim type TASK. This completes
8/27 tasks; tasks 3.2–3.4 remain in progress for this group.

## Task 3.2 evidence

Independent review APPROVED; verification status VERIFIED, claim type TASK.
The fresh reviewer suite passed 1,105 tests with nine skipped in 101.18 seconds.
Global Ruff, formatting (221 files), mypy (108 source files) and diff checks passed.
The reviewer also exercised 24 combinations of order, budget and failed query
families, retaining all 12 declared questions and reconciling coverage totals.

The driver accepts explicit `ranking_mode`, `unit` and `population_complete` while
preserving ordinary defaults. Scope and execution requests are produced together
from the existing ranker's final order. Selected/deferred identities, normalized
evidence, unchanged tier/score, per-question outcomes and per-family completion
travel to the CLI directly. Missing evidence stays unknown. Partial graphs cannot
justify tier-zero pruning or completed coverage. Raw query answers survive a later
timeout even when arbitration cannot complete. A completed outcome denotes query
and arbitration completion within the modeled boundary, not a safety assertion.

`benchmarks/measurements/2026-09-13-ranker-scope-smoke.json` records a rebuilt image
whose backend, driver, contracts and CLI hashes match the source. Real PHP (78 bytes)
and JavaScript (69 bytes) fixtures each selected `risky` over the initially higher
static-ranked `safe`. Both selected questions had tier 4 and score 2.5; execution
outcome IDs matched selection, and each family reported one selected/completed,
one deferred and zero unanswered questions. Each produced one controlled finding.
PHP took 61.57 seconds and JavaScript 51.27 seconds under one 900-second lab budget.
This manually declared entry-region smoke does not validate mapper discovery,
PMPro/NodeGoat replay, capability admission or hook latency.

Nine of 27 tasks are now complete; tasks 3.3–3.4 remain for this group.

## Task 3.3 evidence

Independent review APPROVED after one required repair; verification status VERIFIED,
claim type TASK. The fresh reviewer suite passed 1,128 tests, nine skipped, in
105.95 seconds. Global Ruff, formatting (224 files), mypy (108 source files) and
diff checks passed. The context suite contains 23 focused cases.

The generic driver accepts `change_context` and retains its enriched form. Existing
taint evidence optionally projects full first-party method paths/ranges for entry,
callee, same-file field, known hook and sanitizer context. Decoded snapshot changes
and lexical anchors attach to exact questions without changing ranking weights or
introducing another scope selector. Missing base, locations, empty projections,
dynamic/external/depth dependencies and unsupported configuration/deleted-dependency
projections remain explicit gaps. Missing context prevents tier-zero pruning and
complete-negative outcomes; raw answers remain available.

Review found that broad source/sink/sanitizer text matching could mistakenly exempt
an unknown consumer such as `unknown(req.body)` from missing-context reporting.
The production-query regression reproduced exit 1 with that question incorrectly
pruned as tier zero. The task-local repair restricts the exemption to directly
invoked fact-modeled operations. Existing detector matchers are unchanged. After
repair, consumers of a source, nested sink and sanitizer name all remain unresolved;
a direct modeled sink completes, and the direct sanitizer retains tier zero without
a false context gap.

`benchmarks/measurements/2026-09-13-change-context-smoke.json` retains the final image
and matching backend/driver/regions/snapshot/query hashes, plus both runtime checks.
Real Git histories remove a caller guard while keeping helper code unchanged. PHP
reads 124 source bytes and reaches `helper.php:2:helper` in 63.48 seconds; JavaScript
reads 133 bytes and reaches `helper.js:1:helper` through an ES-module import in
39.63 seconds. Both preserve the dirty checkout/index and remove snapshot scratch.
These are controlled change-to-question/witness checks, not proof that the removed
line enforced security, that the defect is new, or that a hook meets its latency bar.

Ten of 27 tasks are now complete; task 3.4 remains for this group.

## Historical task 3.4 blocker

Claim type TASK; status NOT_VERIFIED. The implementer returned BLOCKED, and an
independent fresh-context Kiro debug investigation returned STOP_FOR_HUMAN with
high confidence: the approved sequence lacks a contributor-scan frontend input
retention prerequisite. There is no task completion or reviewer approval claim.

The work in progress physically filters declared vendor inputs, routes explicit
language partitions through one existing ranker and deadline, and reports input,
excluded/test paths and unsupported coverage. Tests cover stale region paths,
symlink substitution, shared limits, partial census, normalized witness paths and
CLI propagation. Direct JavaScript frontend invocation establishes which engine
is called, but does not override its built-in input filters.

The fresh full suite passed 1,138 tests with nine skipped in 106.05 seconds. Global
Ruff, formatting (227 files), mypy (109 source files), compileall and diff checks
passed. Detection (44/47, zero false positives in 44 controls), split-sink map and
local pair gates passed. The final direct frontend change required adapting one
budget-test runner to recognize its `-o` output argument; the test still verifies
deadline retention. Unit success does not override the failing real graph check.

`benchmarks/measurements/2026-09-13-partition-retention-blocker.json` records exact
packaged source hashes, readable input witnesses, frontend help and failures. The
direct frontend input contained `server/api.js` (51 bytes) and `tests/probe.js`
(41 bytes), with original hashes verified. The graph contained only the server
file and handler. Both autodetection and direct `jssrc2cpg` probes exited 1 on the
retention assertion. The instrument correctly refused to report a passing smoke.
PHP and remaining end-of-scan assertions were not reached in this failing run.

Pinned ASTGen 3.50.1 unconditionally excludes test directories; Joern 4.0.625 also
filters test filename suffixes. Installed/versioned options provide no supported
inclusion override. The upstream sources and concrete repair acceptance checks
are linked in the proposed prerequisite. Do not mark task 3.4 complete by moving
test files or declaring frontend omission acceptable. Its implementation remains
uncommitted pending the foundation repair, review and fresh runtime verification.

The packaged failure preceded a comment-only clarification in `cpg/backend.py`
removing the disproven attribution to legacy frontend autodetection. No executable
backend behavior changed after that measurement.

## Task 3.4 final verification

Status VERIFIED, claim type TASK; independent review APPROVED after two concrete
error-path repairs. The final suite passed 1,151 tests with nine skipped in
100.55 seconds. Ruff, formatting (230 files), mypy (109 source files), compileall,
diff and all three regression gates passed.

The approved foundation repair is committed as `51fabef`. Physically filtered
frontend inputs retain original source paths and bytes; the existing layout facts
mark JavaScript/TypeScript test conventions unshipped. One existing ranker decision
and region/deadline budget cover supported partitions. Browser-path source remains
available, while runtime origin stays unspecified without declarations. Vendor,
cross-language, TypeScript-property and C-bounds limits remain explicit.

Exact source names now travel in taint, dominance and configuration batch censuses.
PHP aggregation preserves them only when every shard provides named evidence.
Synthetic `<unknown>` nodes cannot mask missing real files. Unreadable unknown
extensions produce `source_unreadable` coverage instead of raising PermissionError.
Legacy callbacks that omit census entirely retain their existing compatibility
behavior; production queries emit the named census, and unsupported/partial
sharded evidence cannot establish a complete negative.

`benchmarks/measurements/2026-09-13-partition-scope-smoke.json` records the final
image and ten matching source/query/fact hashes. The real mixed graph retains
seven JavaScript files (296 bytes), including tests and browser source, plus two
PHP files (97 bytes), with vendor nodes absent. One question is selected and three
deferred; unsupported Go is reported. No false empty/unavailable census remains.
PHP and JavaScript change-context controls both complete with unchanged-helper
witnesses, preserving dirty checkout/index state and cleaning snapshot scratch.

A real Python graph supplies named census from all three production query kinds.
Those unchanged Scala sources were verified before the separate PHP transport and
discovery repairs; the artifact records that narrower provenance explicitly. A
real filesystem permission probe verifies the repaired error path, using a clearly
labeled synthetic graph callback.

The mixed smoke took 197.76 seconds under a 900-second lab allowance; PHP/JavaScript
context checks took 117.3/95.9 seconds. These manually declared controls do not
establish automatic scope discovery, production hook latency or alert quality.
Group 4 now owns novelty comparison, actionability admission and compact output.