Add Ollama warmth lifetime scoring as bounded placement tiebreaker

[toasterbook88]

Summary

Promote the resident-model "is loaded" boolean into a continuous 0.0-1.0 warmth score derived from Ollama's /api/ps expires_at and default_keep_alive. Warmth becomes a bounded tiebreaker at position 10 of the placement rank comparator, after RAM, GPU, pressure, and reservation ratio.

FilterCandidates is unchanged: warmth is consulted only among nodes that already passed eval.Eligible(). It cannot promote an undersized node, and the three-bucket discretization (cold/warm/hot at 0.5 and 0.9) keeps ranking stable.

What ships

• ResidentModel.ExpiresAt (time.Time, omitempty) and ResidentModel.WarmthScore (float64, omitempty) in internal/models/types.go
• OllamaInfo.DefaultKeepAlive (string, omitempty) for the process-level Ollama default
• ApplyOllamaWarmth and DefaultOllamaKeepAlive helpers in internal/facts/local.go (exported for testability)
• OllamaDiscoveryScript updated to read ollama ps -qq (JSON path, Ollama 0.3.10+) and fall back to the existing awk parser on older Ollama; queries /api/ps for default_keep_alive and falls back to 5m when missing or unparseable
• modelWarmthRank and warmthToRank in internal/placement/empirical.go (3 buckets: 0 cold, 1 warm, 2 hot at 0.5 and 0.9)
• modelWarmthRank wired into rankKey at position 10 of the comparator in internal/placement/ranker.go
• 11 new tests in internal/placement/warmth_test.go covering: warmth loses to allocatable RAM, warmth breaks ties on equal RAM, warmth is ignored when FilterCandidates rejects, boundary cases (0, 0.5, 0.51, 0.9, 0.91, 1.0, 2.0), highest-relevant-wins, other-runtime warmth is ignored, and time math for zero / future / past ExpiresAt

Safety contract

• Warmth is strictly a tiebreaker. FilterCandidates (ranker.go) calls eval.Eligible() before any ranking begins; warmth cannot override RAM, GPU, or pressure eligibility.
• The 3-bucket discretization (cold/warm/hot) means warmth is stable under small time changes and cannot induce rank flips on transient observations.
• All new fields use omitempty, so older Ollama, other runtimes (llama-server, mlx_lm.server), and zero/expired ExpiresAt all leave WarmthScore at 0 and behave as cold.
• Probe path tolerates ollama ps -qq failing on older Ollama (e.g., 0.23.3 rejects -qq with "unknown shorthand flag") and gracefully falls back to the existing awk parser, which emits no expires_at — those entries remain cold.

Quality gates

• go build ./... ✓
• go test ./... -count=1 ✓
• go test -race ./... -count=1 ✓
• gofmt -l . clean ✓
• go vet ./... clean ✓
• make coverage ✓ (knowledge 90.9% / api 80.9% / mcp 83.7% / ui 94.0% / total 69.1%)
• ./hack/verify-repo-truth.sh ✓
• make build ✓ — binary exposes expires_at and default_keep_alive in axis facts --format json

Test coverage

11 new tests in internal/placement/warmth_test.go:

Test	Locks in
TestRankCandidatesWarmthLosesToAllocatableRAM	small-hot loses to large-cold (warmth cannot promote undersized node)
TestRankCandidatesWarmthBreaksTieOnEqualAllocatableRAM	warmth resolves ties on equal RAM
TestRankCandidatesWarmthFilteredBeforeRanking	hot node with insufficient RAM is dropped by FilterCandidates
TestWarmthToRankBoundaries	strict > comparisons at 0.5 and 0.9 (exactly 0.5 is cold; exactly 0.9 is warm)
TestModelWarmthRankPicksHighestRelevant	ranks by highest matching model's warmth, not average
TestModelWarmthRankIgnoresOtherRuntimes	non-ollama warmth is ignored
TestApplyOllamaWarmthTimeZero	time.Time{} (omitted) → all cold
TestApplyOllamaWarmthInFuturePopulates	future ExpiresAt → WarmthScore > 0
TestApplyOllamaWarmthPastExpiresAtIsCold	past ExpiresAt → cold
TestDefaultOllamaKeepAliveFallbacks	empty/garbage/zero → 5m
TestDefaultOllamaKeepAliveParses	valid durations parse; 5m default when negative

Backward compatibility

• All new fields are omitempty, so older /api/ps payloads and existing fixtures continue to unmarshal cleanly.
• Older Ollama (<0.3.10) falls back to the awk parser; warmth for those nodes is 0 (cold), which is the correct conservative behavior.
• The ranker comparator insert is additive — no existing comparator field is renamed, retyped, or removed.

Notes

• No public-repo-sensitive content (no real hostnames, IPs, SSH users, model names beyond generic placeholders, or per-host output) is included in this PR.
• Author identity is the project's public-safe contributor handle.

[gemini-code-assist]

Code Review

This pull request introduces model warmth scoring and ranking for Ollama resident models. It updates the local discovery script to parse JSON from ollama ps -qq and retrieve the process-level default_keep_alive duration, which is then used to compute a continuous warmth score based on the model's expiration time. This score is integrated into the candidate ranking logic as a bounded tiebreaker. Feedback on the changes includes making the embedded Python script more robust against unexpected JSON structures or types, and handling bare integers representing seconds in DefaultOllamaKeepAlive to prevent parsing failures.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.