QVAC-22567 feat[bc]: remove dynamic tools (tools_compact) from llm-llamacpp addon

[iancris]

🎯 What problem does it solve?

The dynamic tools feature (tools_compact) lets a caller add tool definitions mid-conversation and have them trimmed from the KV cache once the tool-call chain resolves. It only works for Qwen3 — it depends on a custom Qwen3 dynamic template — and on sliding-context anchor bookkeeping that does not carry over to newer LLM models. This PR removes the feature from the llm-llamacpp addon.

Introduced in #706 and #1379 (QVAC-16769). The SDK-side removal (toolsMode) is split into companion PR #3380.

📝 How is it solved?

• Deleted ToolsCompactController and Qwen3ToolsDynamicTemplate (normal Qwen3 templating keeps using getFixedQwen3Template)
• The tools_compact config key is no longer parsed
• Stripped the tools-anchor plumbing from ContextSlider, ContextShifter, ReasoningBlockCompactor, TextLlmContext/MtmdLlmContext, LlamaModel, and ContinuousBatchScheduler — general context sliding, M-RoPE sliding, and reasoning compaction are unchanged
• Removed the two-pass (with/without tools) tokenization, the PromptLayout/validatePromptPolicy prompt-shape validation chain, and the nPastBeforeTools/toolsTrimmed debug stats
• Deleted the feature's unit tests, tools-compact integration + mobile test suites, docs, and examples; adapted shared tests to the new signatures (kept general coverage — e.g. the tail-trim interaction tests were retained and renamed since the branches they pin still exist)
• No version bump — releasing is deferred to the release flow

🧪 How is it tested?

• Full local build (bare-make generate/build/install) and C++ unit tests — 704 ran, 698 passed, 0 failed, 6 pre-existing environmental skips (GPU-gated / fixture-gated); JS lint clean; mobile test manifest regenerated with group coverage validated
• The identical addon content already passed the full labeled CI suite on the pre-split combined branch: cpp-tests on linux/darwin/windows, full prebuild matrix, iOS + Android mobile E2E, desktop integration on darwin-arm64/linux-arm64. The 3 red desktop platforms (test-linux-x64, test-darwin-x64, test-win32-x64) were verified byte-identical to same-day pre-PR baseline runs on other branches (pre-existing, non-gating)
• Repo-wide grep for tools_compact|toolsCompact|tools_at_end|Qwen3ToolsDynamic|nPastBeforeTools|PromptLayout|validatePromptPolicy returns only historical changelog entries

⚠️ Any breaking changes?

Yes — the dynamic tools feature is removed. A lingering tools_compact config key is now accepted but ignored with a deprecation warning (a tombstone in commonParamsParse) for one release, so existing configs — including older SDK builds that still emit the key — keep loading rather than failing model load on an unknown argument. Plan to drop the tombstone (making the key an error) in a future release once no consumer emits it.

BEFORE:

// old — addon config key enabling dynamic tools (Qwen3-only)
const config = { tools: 'true', tools_compact: 'true' }

AFTER:

// new — tool definitions always stay in place; a leftover tools_compact
// key is ignored (deprecation warning) rather than toggling anything
const config = { tools: 'true' }

[github-actions]

Review Status

Current Status: ✅ APPROVED
Approvals so far: Team Lead: 1, Member: 1

[github-actions]

License compliance — clean

No new dependency license findings in this PR.

Warn-only (shadow) mode — this check does not block merges yet.

Updated automatically by the canonical license compliance workflow.

NOTICE presence (advisory)

Missing NOTICE (advisory, does not block):

• ./.github/actions/release-merge-guard
• ./docs/website
• ./packages/ggml-coload-smoke
• ./packages/fabric/test/integration
• ./packages/inference-addon-cpp/mobile
• ./packages/sdk/e2e
• ./packages/llm-llamacpp/benchmarks/performance
• ./packages/llm-llamacpp/benchmarks/server
• ./packages/vla-ggml/sim/server
• ./packages/embed-llamacpp/benchmarks/performance
• ./packages/embed-llamacpp/benchmarks/server
• ./packages/asr-ggml/benchmarks/server

[github-actions]

Mobile integration tests — @qvac/llm-llamacpp (iOS)

Result: passed

metric	value
Devices passed	26
Devices failed	0
Test cases total	78
Test cases passed	78
Test cases failed	0
Test cases skipped	0

View workflow run

[github-actions]

Mobile integration tests — @qvac/llm-llamacpp (Android)

Result: passed

metric	value
Devices passed	21
Devices failed	0
Test cases total	63
Test cases passed	63
Test cases failed	0
Test cases skipped	0

View workflow run

[github-actions]

QVAC E2E — android — ✅ all tests passed (91/103, 2060s)

Config: suite=smoke · filter=(none) · exclude=(none)
Device pool: Google Pixel 9 Pro - Android 15 — ☀️ highly available
View run · Artifacts: reports · Device Farm logs

[donriddo]

isQwen3Architecture fix looks good. Exported from the header, six test cases pin the exact-match boundary including the qwen35 false case that was the original risk. Good to go.

[tobi-legan]

QA review summary: the existing isQwen3Architecture thread is addressed. The main remaining concern is release ordering: this addon PR intentionally removes tools_compact/dynamic tools, but current SDK and inference code still expose toolsMode dynamic and emit tools_compact=true. Before release, the companion SDK/inference PR should remove or loudly deprecate that public mode so users do not get silent behavior changes. The test ask is limited to proving the remaining static tool-calling path, or covering dynamic-mode rejection/deprecation in the companion PR. QIP triage: public addon/SDK contract removal trigger applies; proposal scope is removal and release ordering of dynamic tools.

[tobi-legan]

Re-reviewed the latest commits. Prior comments are addressed: tools_compact=true now rejects clearly, tools_compact=false remains an ignored tombstone, and the warm tool-calling follow-up now validates declared parseable tool calls. No further findings from my pass.