feat(pflash): ee7 early-exit drafter + anchor-transitive cascade + regime router

[easel]

Summary

Adds the anchor-transitive cascade for the qwen3 pflash path (default-on, addresses the 64K NIAH cliff) along with drafter changes that thread the new anchor scan into the existing pflash drafter. Includes a focused C++ unit test for the anchor-transitive behavior and two design/configuration docs.

Files

• server/src/qwen3/anchor_scan.{h,cpp} - new anchor scan implementation
• server/src/qwen3/qwen3_drafter.{h,cpp} - drafter integration of anchor scan (~200 net LOC of changes)
• server/src/server/http_server.h, server/src/server/server_main.cpp - minor wiring touch-ups
• server/test/test_anchor_transitive.cpp - unit test for anchor-transitive cascade
• docs/anchor-transitive.md, docs/pflash-compress-cfg.md - design / config notes

The earlier-claimed ee7 early-exit drafter, regime router, adaptive_keep_ratio, score_range, and the other four C++ unit tests are NOT in this PR's diff after the Phase 1-3 cleanup. The full A/B and prefix-cache experiment writeups previously listed are likewise not part of this scope.

Dependencies

None. After the Phase 1 cleanup this PR is independent and can land on its own (no remaining references to sibling-PR headers such as c2_gate.h or shared scaffolding).

Note: like the other server-layer split PRs, this branch cannot be built standalone because CMake still cross-references files that live on sibling branches; CI build validation has to happen on the integration branch.

Test plan

• test_anchor_transitive passes locally and in CI on the integration branch
• Qwen3.6 pflash run does not regress on the 64K NIAH point that motivated the anchor-transitive default-on change
• Drafter warm-path behavior unchanged for the non-anchor regime (smoke test via a short pflash request)
• Note: server PRs in this split cannot be validated standalone due to CMake cross-references; full build/test must run on the integration branch

🤖 Generated with Claude Code

[cubic-dev-ai]

4 issues found across 9 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="server/src/qwen3/anchor_scan.cpp">

<violation number="1" location="server/src/qwen3/anchor_scan.cpp:27">
P2: Anchor scan incorrectly searches one position when `body_end < ngram`, causing false-positive forced chunks from non-body tokens.</violation>

<violation number="2" location="server/src/qwen3/anchor_scan.cpp:31">
P1: `max_anchor_hits` is effectively capped at 8 due to fixed `hit_pos[8]`, so long-context configs (16/32) do not work as intended.</violation>
</file>

<file name="server/src/qwen3/qwen3_drafter.cpp">

<violation number="1" location="server/src/qwen3/qwen3_drafter.cpp:132">
P2: Deprecation warning is logged on every request in a hot path; it should be emitted once, not per compression call.</violation>
</file>

<file name="docs/anchor-transitive.md">

<violation number="1" location="docs/anchor-transitive.md:13">
P3: Documentation references `bench/2026-05-25_longbench_hotpotqa/` but the `bench/` directory does not exist anywhere in this codebase, so the reference is unresolvable.</violation>
</file>

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[cubic-dev-ai]

P1: max_anchor_hits is effectively capped at 8 due to fixed hit_pos[8], so long-context configs (16/32) do not work as intended.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At server/src/qwen3/anchor_scan.cpp, line 31:

<comment>`max_anchor_hits` is effectively capped at 8 due to fixed `hit_pos[8]`, so long-context configs (16/32) do not work as intended.</comment>

<file context>
@@ -0,0 +1,169 @@
+
+    for (int qi = 0; qi + ngram <= (int)query_pool.size(); ++qi) {
+        int hits = 0;
+        int hit_pos[8];
+        for (int p = 0; p <= search_end && hits <= cfg.max_anchor_hits; ++p) {
+            bool same = true;
</file context>

[cubic-dev-ai]

P2: Anchor scan incorrectly searches one position when body_end < ngram, causing false-positive forced chunks from non-body tokens.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At server/src/qwen3/anchor_scan.cpp, line 27:

<comment>Anchor scan incorrectly searches one position when `body_end < ngram`, causing false-positive forced chunks from non-body tokens.</comment>

<file context>
@@ -0,0 +1,169 @@
+{
+    const int n_chunks = (int)forced.size();
+    const int ngram    = cfg.ngram;
+    const int search_end = std::max(0, body_end - ngram);
+
+    for (int qi = 0; qi + ngram <= (int)query_pool.size(); ++qi) {
</file context>

Suggested change

	const int search_end = std::max(0, body_end - ngram);
	if (body_end < ngram) return;
	const int search_end = body_end - ngram;

[cubic-dev-ai]

P2: Deprecation warning is logged on every request in a hot path; it should be emitted once, not per compression call.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At server/src/qwen3/qwen3_drafter.cpp, line 132:

<comment>Deprecation warning is logged on every request in a hot path; it should be emitted once, not per compression call.</comment>

<file context>
@@ -64,11 +65,122 @@ static int env_int(const char * name, int fallback) {
+        const int nv = env_int("PFLASH_COMPRESS_ANCHOR_NGRAM", -1);
+        const int lv = env_int("DFLASH_COMPRESS_ANCHOR_NGRAM", -1);
+        if (nv >= 0) return nv;
+        if (lv >= 0) { fprintf(stderr, "[WARN] DFLASH_COMPRESS_ANCHOR_NGRAM deprecated, use PFLASH_COMPRESS_ANCHOR_NGRAM\n"); return lv; }
+        return 4;
+    }();
</file context>

[cubic-dev-ai]

P3: Documentation references bench/2026-05-25_longbench_hotpotqa/ but the bench/ directory does not exist anywhere in this codebase, so the reference is unresolvable.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At docs/anchor-transitive.md, line 13:

<comment>Documentation references `bench/2026-05-25_longbench_hotpotqa/` but the `bench/` directory does not exist anywhere in this codebase, so the reference is unresolvable.</comment>

<file context>
@@ -0,0 +1,15 @@
+
+Empirical result: F1=0.628 on LongBench HotpotQA at ee7 + keep=0.15
+(vs uncompressed F1=0.697). This is the ceiling for attention-score-based
+prefill compression on this task; see bench/2026-05-25_longbench_hotpotqa/.
+
+On by default. Disable via PFLASH_COMPRESS_ANCHOR_TRANSITIVE=0.
</file context>

[easel]

Closing this PR — it duplicates open PR #274 (feat/pflash-drafter-ee7) by @dusterbloom.

The substantive content here — anchor_scan.{cpp,h}, the qwen3_drafter changes, test_anchor_transitive.cpp, anchor-transitive.md, and pflash-compress-cfg.md — is byte-identical to what's already in #274. The original commit d4546a51 (dusterbloom, 2026-05-27) predates my squash 36979404 (2026-05-31), so attribution belongs upstream.

Reviewers: please land #274 as the canonical source.

Apologies for the inadvertent re-attribution — this should never have been opened against the original work. If any small subtractive cleanup is warranted (e.g., removing the regime_router wiring), I'll open that as a separate, narrowly-scoped PR with honest authorship after #274 lands.