Open_Ultra_SAST/.kiro/specs/pre-push-safety-net/status.md

Pre-push safety net status — v1.2.0-alpha.1

Updated 2026-09-16. Phase: implementation. Requirements, design and tasks are approved. 28/29 executable tasks complete; task 8.3 remains blocked. Rollout: NO-GO. The alpha tag publishes experimental implementation and evidence; it does not approve normal hook capabilities or production enforcement.

Next release milestone

M1a is met and the M1b comparison now works through the production rules. The other four milestones remain unverified, and rollout remains NO-GO.

Every veto M1a recorded has been repaired. On 2026-09-16 the production comparison classified the pinned NodeGoat revisions without any lifting: the vulnerable revision is new with the three eval witnesses attributed to the change that introduced them, the fixed twin is analyzed and quiet, and the unchanged revision is unchanged. Production and the recorded view agree exactly, which is how we know the vetoes were repaired rather than bypassed. No capability is enabled, no defect is admitted and no alert is emitted. See M1a and M1b evidence.

The exit run is recorded. Every veto M1a listed is repaired, plus two the repairs exposed: the comparison's own degradation and boundary checks, and the runner handing every candidate the whole transaction's gaps. On the vulnerable revision admission now rejects on capability_unavailable alone.

One half of M1b remains. The milestone also asks for a developer-readable finding carrying a supported security consequence and a concrete repair direction. Both are owned by an evaluated capability declaration and the registry is deliberately empty, so nothing renders them. Writing that text without a declaration is what admission exists to prevent, so it is not worked around. Task 10.10 renders it through an explicitly experimental declaration with eligibility still disabled.

M1a ran on 2026-09-16 within its one-week box. The raw <lambda>2 witnesses are present at contributions.js lines 32–34, the fixed twin was genuinely analyzed and quiet, and the unchanged revision was not reported as a new regression, so the kill criterion did not trigger. Five vetoes fired and were recorded; with all five lifted the detection still stops at operation_correspondence_unresolved, because a newly introduced operation sits on a changed line and has no base anchor. That fifth veto was not in the earlier diagnosis and is now M1b's central repair. See M1a evidence.

The 28/29 task count is implementation progress, not proximity to rollout. Follow the accepted release milestone spec: usefulness → speed → shared PHP/Python transfer → independent qualification → narrow advisory rollout.

Security capability, 2026-09-17

Group 12 closed three defects that limited what the tool can establish, each measured on the pinned NodeGoat revisions with detections unchanged throughout.

• Access control could not answer anywhere. Dominance and configuration were never asked for change context, because the evidence pass dispatches taint only. Each family now describes its own scope. Head completion rose from 205 to 298 of 436, context_projection_unavailable fell to zero, and access control went from 0 to 52 completed questions.
• Admission rejected genuine findings. The operation symbol was the text before the first parenthesis, so const preTax = eval(...) was the operation const preTax = eval, which no qualified declaration could ever name. It is now the trailing qualified callee, with receiver spelling preserved.
• Unresolved questions said why, and then there were none. The remaining 138 each named a function its file does not define. A route-registration module registers handlers other modules define, and regions were scoped to methods those files do not hold. The entry record keeps the handler name, because the authorization question is cross-file by nature, and the region is now scoped to the file instead. Every question the scan asks now completes on both revisions, 299 of 299, with detections unchanged.

Evidence: group 12 final and region attribution.

Verified

Immutable pushed snapshots, shared multi-ref deadlines, ranker-owned scope, physical vendor exclusion, compatible immutable graph/result reuse, actionability admission, compact truthful reporting and non-destructive hook integration are implemented. Advisory and blocking share the same evidence bar; optional model output cannot bypass it.

Contributor-scan task 2.19 is verified: the pinned frontend now retains the original Gruntfile, bringing the real NodeGoat census from 43/44 to 44/44 files. Partial, absent and malformed census no longer masquerade as a complete result or a falsely empty graph. The final code verification recorded 1321 passed tests, nine skipped, passing canonical gates and a packaged native PHP/JavaScript partition smoke. See owner verification.

Rollout blockers

1. Useful checks within the deadline are unproven. The fresh post-census profile completed all 21 measurements: all 18 changed/cold/growth/multi-ref transactions timed out; the three identical-tip replays took 18.521–18.874s with valid graph/query reuse. Still 0/24 declared target checks completed. Maximum observed cancellation/report overrun was 0.874s, below the two-second allowance. After M1 establishes the useful comparison, amortize engine startup/loading and rerun the frozen seven-class profile; require warm p95 <=30s and >=95% completed supported checks.
2. Target and change-context evidence remains incomplete. M1a measured exactly which contracts fail on the Node case: per-family change context, question-owned vendor relevance, operation correspondence for changed lines, file-wide correspondence ambiguity and query-to-operation provenance. VAmPI preserves raw findings at all three known sites, but the exact SQLI function question/transitive witness and supported authorization context are unresolved. Raw findings or rank positions cannot establish an actionable new regression. Repair this in the shared scan boundary and verify the actual associated query/path evidence.
3. Independent quality qualification is missing. Freeze and review untouched Node API vulnerable/fixed/benign cases, including middleware/module/async context. Ghost remains reserved with missing reviewed inputs. PHP and Python need their own independent language/framework/family populations; Node results cannot qualify them. Require >=95% observed actionable precision with counts/intervals, >=90% supported recall, nonzero reviewed emitted alerts/positive controls and zero known fixed/benign false alerts.
4. No current capability is eligible. The earlier registry is NO-GO and becomes stale under the repaired runtime/core identities. Loading it enables zero capabilities. After the preceding evidence passes together, regenerate exact-semantic eligibility and rerun packaged acceptance validation. Do not relabel old measurements as current.

C/libpng arithmetic and bounds remain explicitly unsupported; this is not a C/C++ detection qualification. Advisory rollout cannot be justified by silence from an unqualified registry.

Ownership and revalidation

Engine lifecycle, source/census and context/witness normalization belong to contributor-scan. The hook consumes those contracts; no second engine, source IR, ranker priority exceptions or differential graph mutation is introduced here. Graph mutation would need a separate cold-equivalence experiment.

After owner fixes, refreeze identities and rerun ordinary scan gates, Python regressions, PHP-to-Node transfer, the representative cache/deadline profile, independent per-capability quality evaluation and packaged hook validation. Keep task 8.3 unchecked and the feature in implementation until those mandatory results are established.

Evidence: group 8 validation, post-census runtime scorecard, census repair.

The owner investigation now distinguishes missing family context producers, globally propagated vendor/correspondence gaps, and query-to-operation identity. See remediation research. These are next implementation prerequisites, not completed fixes or changed acceptance gates.

An isolated Joern session probe matched the full Node census with four warm requests in 0.221–0.467s and verified cancellation/reaping in 0.043s. It identifies a promising engine lifecycle change; full-query equivalence and changed-code performance remain unverified.

Family detection census, 2026-09-19

Four of seven declared families establish findings on the two development subjects: injection, access_control, config_secrets and untrusted_destination. Output encoding, path, prototype and deserialization have established nothing anywhere. Every question completed on both subjects, 299 of 299 on NodeGoat and 35 of 35 on VAmPI, so the silences are answers rather than failures — but three of them are diagnosed gaps rather than clean code: NodeGoat's documented SSRF is invisible because needle.get is not in the sink vocabulary, its XSS is invisible because output encoding declares only DOM sinks, and its two access-control defects raise no obligation because the modelled data operations live in DAO files while the family's scope is the handler's own file.

The census also found and fixed one false positive: NodeGoat's only prototype report was the word set in a code comment, matched because the sink text clause searched a whole assignment node. Nothing about this changes the release gates. No capability is qualified by these subjects.

Two of the three gaps are closed. The untrusted-destination vocabulary now holds the client surface a Node server actually calls, and NodeGoat's SSRF is reported at app/routes/research.js:16 with a witness naming the call. Output encoding now asks about the server half of its own family and establishes a finding at app/routes/research.js:25, and it states in its declaration that template-mediated escaping is not something it can decide. Five of seven families now establish findings.

Access control was the third gap and its structural half is closed. The obligation now follows the one named call a handler writes into its data-access module, attributed by the graph's own provenance rather than by name, and located at the call a contributor changed. On NodeGoat it is raised at six handlers where it was raised nowhere, and three reach a verdict. All three are corroborated and none is reported without a judge, which is the honest rung: entailing a carried obligation reported a signup handler that is public by design, because its authenticated siblings share the same data module. VAmPI is unchanged.

The commonest collection read now carries the obligation too, so NodeGoat's documented insecure direct object reference raises one where it never had. find had been left out because the name is shared with array searches and jQuery selectors, so it is claimed only in the argument shape of a query document. Measured on a plugin that ships jQuery, all 88 of its find calls stay unclaimed. Seven NodeGoat handlers now reach a verdict, all corroborated; the census is unchanged, because the corroborated band needs a judge it does not configure.

What a session read discharges is now decided by where its value goes. A destructuring spells session.userId away, so the arbiter walks back from the names the query is given; two NodeGoat handlers that were correct code stop being asked about, while the memo page that reads every memo and merely renders the session's user stays reported. The sibling scope was decided by measurement and left at the file: on this subject the file scope entails two handlers, both public routes and both wrong, and an application-wide scope entails four true defects and the same two wrong ones. Neither scope can tell a public route from a protected one, so the carried rung stays capped and that question stays with the judge.

Configuration reports every setting it decides, not the first one in the file. NodeGoat's server.js now carries both its permissive session and its disabled template auto-escaping, each with its own site; the escaping defect had been detected on every run since task 14.3 and reported on none, because an earlier setting in the same function won the single verdict.

The PHP record is corrected. The malformed-CPG bug this project reported upstream, a global inside a closure giving a node two AST parents, was fixed in Joern 4.0.625 and the image has pinned that version throughout; it is verified today on the real trigger file, on 4.0.625 and 4.0.631 alike. What remains under task 13.2 is runtime alone: a 2,526-question batch over a 673-file plugin does not finish in thirty minutes, and a six-fold larger ceiling only made the failure take longer. The engine pin stays where it is, because the newer release measures the same here.

That runtime reading was itself incomplete, and the correction matters more than the reading. PHP is not silent: at 25 and 100 regions PMPro completes with no timeout at all and establishes 19 and 43 findings, including a true-positive SQL injection at adminpages/discountcodes.php:749 of exactly the class the PHP fact table predicted. It looked silent because the census harness ordered its verdict on question completion before findings, and because only the 500-region runs had been examined. The harness is fixed. What remains is a scale limit of roughly one to two seconds per question, and one real gap that is bookkeeping rather than detection: no PHP question is ever marked completed, even in the runs that find things.

Closing the second gap surfaced four defects, three of which were false positives it would otherwise have shipped: a permissive-value set shared across settings whose booleans have opposite polarity, a dotted setting matched on its trailing segment alone, a flow query asked of a call's receiver as well as its arguments, and a file's settings ordered by the spelling of their line numbers. Template auto-escaping disabled at server.js:135 is detected and still not reported, because the arbiter returns one verdict per question and that file has an earlier permissive setting; task 14.6 carries it.

Evidence: family detection census, untrusted destination vocabulary, output encoding family, access control across the module split, collection read obligation, sibling scope and identity, every permissive setting, PHP runtime after the graph fix, PHP detection corrected.

Strategy, 2026-09-21

Sequence amended after the family census and the PHP correction: close M3 at a completing scale first (task 13.3), fix PHP question completion because admission reads outcomes (13.4), reserve the untouched M4 population in parallel starting now (15.1), and give the remaining engineering to M2 graph reuse, which cannot be scoped away because meeting the latency gate by analysing less is explicitly not the same result. The v1 advisory scope is declared as the families whose claims can be admitted at the entailed rung (15.2): injection, untrusted destination and configuration. Access control is excluded because a carried obligation corroborates by design and would ship as silence.

No acceptance threshold and no owner boundary changes. The risk that remains is that every number so far comes from subjects used to build the analyser, and one week of examining two of them surfaced seven false positives; the first independent evaluation should be expected to fail.

PHP reaches its declared target, 2026-09-22

The M3 PHP target is detected: classes/class.memberorder.php:936:getMemberOrderByCode, a request value concatenated into SQL with no prepare. It had never been reached, and nothing about detection changed to reach it. The population became affordable.

Four repairs did that, and three of them are the same rule the design now carries. Tier-zero pruning was switched off for a whole scan by any degradation, and a multi-language repository always carries one. Batches were sized by request count when the cost is per matched sink, so a family whose sinks are echo and print was priced like one whose sinks are rare. A per-query timeout cancelled the entire scan, abandoning work with half an hour of budget left. And nothing was held back for arbitration, so a scan could spend everything on queries and report none of what they answered.

On the 500-region population that previously completed zero questions at every budget tried: 271 of 406 completed, 213 findings, 2,950 seconds. The three declared comparisons are running at that budget; detection is not transfer, and only the comparison decides M3.