spore.host integration tracker (standing): truffle/lagotto/spawn deltas + acquisition seam

[scttfrdmn]

Tracks integration with the spore.host libraries and the upstream deltas surfaced while reading their real API (spec §5/§17: read from source, log deltas, don't invent signatures). Scott can patch these repos same-day.

Upstream issues filed

• truffle Card name "RTX PRO 6000" resolves only via single-token alias; empty match falls back to .* truffle#90 — "RTX PRO 6000" resolves only via single-token alias; empty match falls back to .*.
• spawn Confirm acquisition seam: spawn owns RunInstances (no pre-acquired-instance bring-up); no exported ECR/container primitive spawn#351 — spawn owns RunInstances (no pre-acquired-instance bring-up); no exported ECR/container primitive; LaunchResult has no timestamp/region.
• lagotto Blessed library path for blocking single-target acquire? No Snipe()->handle; poller needs DynamoDB + caller-owned retry lagotto#73 — no blocking Snipe()->handle; poller needs DynamoDB + caller-owned retry loop.

Net architectural finding (reshapes spec §5)

Spec assumed: lagotto acquires -> spawn brings up on the live instance. Real spore.host model is the inverse: spawn owns the RunInstances launch (the launch is the capacity test); lagotto is the retry policy above it. For calque's block-and-wait, the leaner path is likely spawn.launcher.Provision + a retry loop keyed on watcher.ClassifyFailure (no DynamoDB dependency) — pending the owners' answer on lagotto#73. Blocked design decisions live here until those resolve.

[scttfrdmn]

Seam RESOLVED by spore.host owner (all 3 upstream issues answered)

• spawn#351: seam confirmed inverted from spec §5. Build against launcher.Provision(ctx, *aws.Client, aws.LaunchConfig, Options) (*aws.LaunchResult, error) as the acquire+bring-up primitive — spawn owns RunInstances; there is no BringUp(handle). Retry is the caller's loop keyed on LaunchError.Code (pkg/aws/client.go:283, verbatim AWS code — classify without string-matching). LaunchResult Region+LaunchTime addition tracked in spawn#352; headless container-run primitive in spawn#353. Now v0.73.0.
• lagotto#73: blessed path for block-and-wait single-target is (b) — Provision + retry keyed on watcher.ClassifyFailure, NOT the DynamoDB-backed poller (poller drags in DynamoDB/Substrate for something fundamentally stateless). A stateless watcher.Snipe(ctx, target, opts) (*MatchResult, error) wrapper is coming; scoped against spawn#352.
• truffle#90: .* match-all fallback confirmed a footgun; root cause is a 3-word token cap vs a 4-word canonical key. Owner will add "rtx pro 6000" alias + a ResolveCard(card) ([]string, error) returning ErrNoMatch.

calque's build decision (leanest correct path)

Import spawn only for the spike. Behind a calque-side plan.Acquirer seam:

• Provision + own retry loop keyed on LaunchError.Code (capacity codes = wait, terminal = fail).
• When watcher.Snipe lands, swap it in behind Acquirer — no plumbing change.
• Avoids lagotto's DynamoDB dependency entirely for block-and-wait.

Result collection uses LaunchConfig.PreStop (aws s3 sync /results s3://… on terminate) + OnComplete. truffle called via ParseQuery→ResolveGPUInstances, guarding the .* fallback until ResolveCard ships.

This issue can close when spawn#352 + the truffle/lagotto helpers merge; interim path is unblocked NOW.

[scttfrdmn]

Real-run infra provisioned (us-west-2), N=1/N~100 authorized

Payload decision (my call, per authorization): Qwen/Qwen2.5-1.5B-Instruct — ungated, ~3GB, fast warm load, real vLLM inference so map_batch stays shape-real. Base image: vllm/vllm-openai + thin warmd/runner/occupancy layer.

Infra created (account-specific identifiers omitted from this public record, per #1):

• S3: a dedicated calque-spike-* bucket (us-west-2) for weights + results.
• ECR: a calque-worker repo (us-west-2).
• Instance IAM: a narrow inline policy on the existing spored-instance-role (reusing spored-instance-profile that spawn knows). Scoped to the calque-spike bucket (S3 rw) + calque-worker repo (ECR pull); only ecr:GetAuthorizationToken is * per AWS requirement. Chose inline-on-shared-role over a new role per direction; kept least-privilege (two guardrail catches during setup corrected an over-broad ECR Resource:* on pull actions -> now repo-scoped).

Next: build the real exec path (S3 sink, on-instance warmd entrypoint, docker build+push, acquire+collect), then launch N=1. Spend gated to N=1 and N~100; will check in before N=100k.

[scttfrdmn]

Capacity investigation: lagotto/spawn/calque EXONERATED — g7/g7e on-demand is held back

Scott (AWS insider) suspected lagotto wasn't working and that us-west-2 az1/2/3 should have on-demand g7/g7e. Tested directly, bypassing calque entirely with raw aws ec2 run-instances:

AZ name→ID mapping (this account, us-west-2): straight-through, no shuffle — 2a=usw2-az1, 2b=az2, 2c=az3, 2d=az4. g7e offered in all 4 AZ IDs; default subnets exist in az1/az2/az3 only (az4/2d has none — but Scott says az4 is full anyway).

Direct raw-CLI launches (no calque/lagotto/spawn in the path):

request	result
g7.2xlarge us-west-2a / 2b / 2c (pinned)	❌ InsufficientInstanceCapacity
g7.2xlarge EC2-chooses-AZ (no pin)	❌ InsufficientInstanceCapacity
g7e.2xlarge usw2-az1 / az2 / az3 (the AZs Scott expected capacity in)	❌ InsufficientInstanceCapacity (all three)
dry-run (config validity), all AZs	✅ succeeds — instance-type + AMI + subnet all valid
quota	768 vCPUs (not blocked); 0 running G instances

Conclusion: calque's Acquirer (via spawn.Provision) returns the EXACT same InsufficientInstanceCapacity that a bare AWS CLI call gets, for a correctly-formed request, with EC2 choosing the AZ. lagotto/spawn/calque are behaving correctly — they surface real scarcity, they don't cause it. My AZ-sweep covers az1/2/3 (the subnet-having AZs = the ones expected to have capacity); az4 exclusion costs nothing since az4 is full.

The gap is fleet inventory vs. on-demand allocation: the GPUs physically exist (Scott's internal view shows them) but are held back from the on-demand shared pool for this account right now — not grantable via RunInstances regardless of caller. The realistic path to a true RTX PRO 6000 K is a Capacity Block / ODCR, or a window when on-demand opens.

Net for the spike: the measured K stands on g6/L4 (K≈73) and g6e/L40S (the closest grantable card, full 1/100/1000 ramp). A 12h parallel snipe matrix (g7+g7e × us-east-1+us-west-2, $0 until landing) is running to catch any brief on-demand opening; calque session --instance g7e.2xlarge produces the RTX PRO 6000 number with zero code changes the moment capacity is grantable.

[scttfrdmn]

Kept OPEN as the standing spore.host integration tracker. Still-live threads (spawn headless-container/ECR primitive → spawn#351/#353, flagged at internal/exec/bootstrap.go:36; --detach stubbed) are detailed in #50 (M10/S4). The original build milestones M0–M4 are closed as shipped.