[reference] rxjava-3.0 iteration-2 regen against skill-async-iteration-2

[jordan-wong]

[reference] rxjava-3.0 iteration-2 regen against skill-async-iteration-2

Reference output from the toolkit's automated instrumentation-authoring
workflow, run against the iteration-2 skill draft (PR #11990, 13 new rules
encoded on top of master).

Not intended to be merged as-is.

Research context

Iteration-2 of the async research cycle in apm-instrumentation-toolkit.
Partner PR: #11974 (reactor-core-3.1 iter-2, all 6 gaps CLOSED).

Rxjava3 iteration-1 (#11939) surfaced 3 Rule #2 divergences (RI-1 super
alias dropped, RI-2 muzzle fail block missing, RI-3 reactive-streams
version regression) plus 4 reviewer-derived rules from @ygree (NR-1..NR-4).
All 7 were encoded into the iteration-2 skill (#11990). This PR is the
regeneration against that updated skill.

Score against iteration-1 gaps: 2 of 7 CLOSED

Ref	Gap	iter-2 result
RI-1	super("rxjava", "rxjava-3") alias	CLOSED ✅ — restored verbatim
NR-1	helperClassNames() array ordering	CLOSED ✅ — matches master
RI-2	muzzle { fail { rxjava2-must-not-match } } block	STILL-OPEN ❌ — skill has rule, model did not apply
RI-3	reactive-streams compileOnly 1.0.3	STILL-OPEN ❌ — regenerated as 1.0.0
NR-2	contextStore() hoist Context.class.getName()	STILL-OPEN ❌ — inlined 5x instead of hoisted
NR-3	No inline //-comments in Advice methods	STILL-OPEN ❌ — advice bodies have narrative comments
NR-4	No banner separator comments in test files	STILL-OPEN ❌ — // --- banners still present

The primary finding — rule-adherence gap

The 5 STILL-OPEN gaps are the primary research signal of this PR. All 5
rules exist in the iteration-2 skill draft (#11990) with explicit ❌/✅
examples and source citations. The model read them and generated output
anyway that violates each one.

This is the same problem class the HTTP cycle diagnosed as N-HTTP-2 earlier
this month: "encoding is not the same as adherence." Adding more prose
guidance to the skill produces inconsistent results — the model applies
some rules naturally (RI-1, NR-1) and doesn't apply others (RI-2, RI-3,
NR-2, NR-3, NR-4).

Contrast with reactor-core-3.1 iter-2 (#11974): all 6 gaps CLOSED. The
reactor result shows the rules work when the model naturally applies them;
this rxjava3 result surfaces the residual adherence gap.

Skill-update prescription for #11990 reviewers: consider whether the 5
open rules belong in reviewer_checks/dd_trace_java/ (deterministic
grep-based enforcement, forces a fix-loop) rather than in prose advisories.
All 5 are mechanically grep-able:

• fail { name = "..." } presence check
• compileOnly ... version: diff vs master
• contextStore() body grep for repeated Class.getName()
• @Advice.OnMethod* body grep for //-comment lines
• Test-file grep for // --- banner patterns

Cost note

$55.01 for this iter-2 run vs $43.23 for iter-1 — MORE expensive despite
partial-fix output. Root cause: the reviewer step looped because the
toolkit's lint adapter looked for the wrong directory
(dd-java-agent/instrumentation/rxjava3-generated/, doesn't exist) and
could not auto-fix any issues. That's a separate toolkit adapter bug,
unrelated to the skill iteration.

Local verification

:check :muzzle :latestDepTest all green after generation.

Cross-references

• Skill under test: skill(apm-integrations): async cycle iteration-2 — 13 rules from #11939/#11940 review feedback #11990 (skill(apm-integrations): async cycle iteration-2 — 13 rules from #11939/#11940 review feedback)
• iter-1 reference: rxjava-3.0 async blind regeneration for #11927 #11939 (rxjava-3.0 iteration-1)
• Sibling iter-2 (reactor): [reference] reactor-core-3.1 iteration-2 regen against skill-async-iteration-2 #11974 (all 6 gaps CLOSED — the "before/after" partner)
• Cycle report: docs/eval-research/runs/rxjava3/attempt2/outcome.md in apm-instrumentation-toolkit
• Skill provenance: docs/eval-research/runs/rxjava3/attempt2/skill-provenance.md

🤖 Generated with APM Instrumentation Toolkit

[dd-octo-sts]

🟢 Java Benchmark SLOs — All performance SLOs passed

Suite	Status
Startup	🟢 pass

SLO thresholds are defined here based on automatically generated metrics. A warning is raised when results are within 5% of the threshold.

PR vs. master results

Scenario	Candidate	master	Δ (95% CI of mean)
startup:insecure-bank:iast:Agent	14.09 s	14.01 s	[-0.4%; +1.5%] (no difference)
startup:insecure-bank:tracing:Agent	12.93 s	12.99 s	[-1.4%; +0.6%] (no difference)
startup:petclinic:appsec:Agent	16.24 s	16.88 s	[-8.0%; +0.4%] (no difference)
startup:petclinic:iast:Agent	16.81 s	16.96 s	[-1.7%; -0.1%] (maybe better)
startup:petclinic:profiling:Agent	16.75 s	16.74 s	[-1.0%; +1.2%] (no difference)
startup:petclinic:sca:Agent	16.91 s	16.80 s	[-0.2%; +1.5%] (no difference)
startup:petclinic:tracing:Agent	16.14 s	16.18 s	[-1.4%; +0.8%] (no difference)

Commit: e0722d5e · CI Pipeline · Benchmarking Platform UI

---

Load and DaCapo benchmarks can be triggered manually in the GitLab pipeline. Results will appear in the Benchmarking Platform UI after completion.