CF-7 Phase E: retrospective grading + new NoKvSlot storm on Qwen3.5 (§7e, open)

[hardcoreerik]

Summary

• Retrospectively applies the Remediation Phase 2/3 Graded capability gate by hand to Meta-Llama's and qwen2.5-coder's already-completed runs (no re-run needed — Phases C/D only changed scoring/gate logic, not B0-B3 answer generation).
• Documents a severe, previously-undocumented NoKvSlot storm hit by two fresh Qwen3.5-9B full 120-question runs (both quants) on fully-fixed master — neither the SeqMax fix (§7b) nor the thinking-suppression fix (§7c) prevents it. This contradicts §7b's earlier "zero NoKvSlot, verified live" claim on the same code.
• Both Qwen3.5 runs' B3 scores are excluded from Phase E results as corrupted-run artifacts, not capability measurements. Root cause is still open (new §7e section, unresolved).

Findings

• Meta-Llama-3.1-8B: Graded capability gate would PASS (B3 58 vs. best baseline B2 46), but ReadyForExpansion still fails overall on segment_terminal_coverage (85.2%, blocking, target 1.0).
• qwen2.5-coder-7b: Graded capability gate FAILS (B3 29 vs. B2 52) — same NO-GO the old all-or-nothing gate reached, now for the documented right reason.
• Qwen3.5-9B-Q8_0 (fresh full run): 96 NoKvSlot incidents, permanent break on the Reviewer role after ~168 clean prior conversations, never recovers. B3 14/120 (worse than B2's 52/120) — void.
• Qwen3.5-9B-Q4_K_M (diagnostic re-run, smaller quant = more VRAM headroom): 59 incidents, same onset point, but partial recovery instead of permanent breakage — evidence pointing toward (not proving) VRAM fragmentation as a contributing factor. B3 40/120 (vs. B2 57/120) — also void.

Test plan

• Both runs' raw logs grepped for NoKvSlot/RECYCLING patterns and cross-checked against the gate JSON verdicts before writing any of this up.
• No code changes in this PR — documentation only.
• grok-review.ps1 adversarial pass
• CodeRabbit review, apply blocker/minor fixes

🤖 Generated with Claude Code

Summary by CodeRabbit

• Documentation
  • Recorded additional Qwen3.5 CF-7 runs invalidated by a severe NoKvSlot storm on the Reviewer role.
  • Added a Phase E retrospective, revising prior Qwen3.5 claims and documenting “graded capability” outcomes.
  • Expanded the bug history with rerun results (including severity progression), open root-cause notes, and guidance to verify raw logs.
  • Updated compatibility notes: Qwen3.5 remains unsuitable for valid full-scale scoring due to unresolved NoKvSlot and related runtime gating issues.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: d554bbd9-f8a0-45dd-8db7-da645c1c694b

📥 Commits

Reviewing files that changed from the base of the PR and between 8b8ec78 and 92207cb.

📒 Files selected for processing (1)

• docs/CF_TEST_RESULTS.md

🚧 Files skipped from review as they are similar to previous changes (1)

• docs/CF_TEST_RESULTS.md

---

📝 Walkthrough

Walkthrough

The documentation records invalid Qwen3.5 CF-7 runs caused by recurring NoKvSlot storms, revises compatibility and bug-history conclusions, and adds Phase E gate verdicts for existing model results.

Changes

Phase E findings

Layer / File(s)	Summary
Qwen3.5 NoKvSlot evidence docs/CF_TEST_RESULTS.md, docs/CONTEXT_FABRIC_BUG_HISTORY.md, docs/CONTEXT_FABRIC_INFRASTRUCTURE_NOTES.md	Qwen3.5 Q8_0 and Q4_K_M runs are marked invalid after NoKvSlot storms; related documentation records incident counts, unresolved behavior, and the absence of a valid full-scale score.
Phase E gate retrospective docs/CF_TEST_RESULTS.md	Existing B0–B3 results receive manual Phase E gating: Qwen2.5 Coder fails Graded capability, while Meta-Llama passes that gate but remains overall NO-GO.

Estimated code review effort: 2 (Simple) | ~10 minutes

Possibly related PRs

• hardcoreerik/TheOrc#57: Restructures the CF-7 bug-history and infrastructure documentation updated here.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title accurately summarizes the main docs changes: retrospective grading and the new Qwen3.5 NoKvSlot storm section.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch docs/cf7-phase-e-retrospective-and-nokvslot-storm

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@docs/CF_TEST_RESULTS.md`:
- Around line 223-227: Correct the ReadyForExpansion explanation so the 0.9
threshold is attributed only to citation_precision, not
boundary_stitch_pass_rate. In the paragraph around the ReadyForExpansion
conclusion, state that citation precision passes while boundary-stitch remains
unverified due to the earlier 0/2 results and lack of re-verification.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 5300215a-ce2d-496f-a2b9-0530ea5273bd

📥 Commits

Reviewing files that changed from the base of the PR and between d43ac96 and 8b8ec78.

📒 Files selected for processing (3)

• docs/CF_TEST_RESULTS.md
• docs/CONTEXT_FABRIC_BUG_HISTORY.md
• docs/CONTEXT_FABRIC_INFRASTRUCTURE_NOTES.md

[hardcoreerik]

@coderabbitai review

[coderabbitai]

✅ Action performed

Review finished.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.

[hardcoreerik]

@coderabbitai review

[coderabbitai]

✅ Action performed

Review finished.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.