verify(P05): 4-layer PASS — REQ-083

---ci---
project: orca
phase: P05
milestone: v0.9
status: verify
---/ci---

.ciagent/CHECKPOINT.json

@@ -1 +1 @@
-{ "phase": "P03/P04/P08", "stage": "verify", "milestone": "v0.9", "phase_role": "execution", "updated_at": "2026-08-05T04:05:00Z", "milestone_complete": false, "verify": { "build": "pass", "go_test": "22/22", "bats": "20/20", "gofmt": "clean", "verify_reqs": "90 consistent" } }
+{ "phase": "P05", "stage": "verify", "milestone": "v0.9", "phase_role": "execution", "updated_at": "2026-08-05T04:15:00Z", "milestone_complete": false, "verify": { "build": "pass", "go_test": "23/23", "bats": "20/20", "gofmt": "clean", "verify_reqs": "90 consistent" } }

internal/scheduler/cel.go

@@ -0,0 +1,536 @@
+// Package scheduler — cel.go implements a minimal CEL-subset evaluator
+// for the CLI-side scheduler constraint expressions (REQ-083, P05).
+//
+// The full CEL specification (google.golang.org/genproto/...
+// googleapis/api/expr/v1alpha1) is intentionally NOT a dependency of
+// this module (see go.mod): adding it for a single callsite would pull
+// in a large transitive graph and contradict the "stdlib + minimal
+// deps" guardrail. Instead this file implements a hand-rolled
+// recursive-descent evaluator for the subset the PRD exercises:
+//
+//   - attribute access on a `node.<name>` object (hostname, kind,
+//     cpus, memory, tags, runtimes)
+//   - string and integer literals (double-quoted)
+//   - comparison operators: == != >= <= > <
+//   - membership: <expr> in <expr>, <expr> not in <expr>
+//   - boolean composition: and, or, not (parenthesised)
+//
+// Anything outside this subset returns an error rather than a silent
+// wrong answer; that is the documented limitation. The grammar is
+// small enough to be unambiguous with a top-down precedence-climbing
+// parser.
+package scheduler
+
+import (
+	"fmt"
+	"strconv"
+	"strings"
+	"unicode"
+)
+
+// EvaluateConstraint evaluates a single CEL-subset expression against
+// the supplied NodeInfo. Returns (matched, err). An expression that
+// references an unknown attribute, uses an unsupported operator, or
+// fails to parse yields an error. Schedule treats a constraint
+// evaluation error as a non-fit (the node is silently skipped) rather
+// than a hard fail because operators routinely write exploratory
+// constraints against attributes the local cluster does not expose.
+func EvaluateConstraint(expr string, node NodeInfo) (bool, error) {
+	p := newParser(strings.TrimSpace(expr), node)
+	if p.len() == 0 {
+		return false, fmt.Errorf("cel: empty expression")
+	}
+	v, err := p.parseExpr()
+	if err != nil {
+		return false, err
+	}
+	if p.tok.kind != tokEOF {
+		return false, fmt.Errorf("cel: trailing input near %q", p.tok.text)
+	}
+	b, ok := v.(bool)
+	if !ok {
+		return false, fmt.Errorf("cel: expression did not evaluate to bool (got %T)", v)
+	}
+	return b, nil
+}
+
+// EvaluateAll returns true iff every constraint evaluates to true
+// against the node (logical AND). An empty constraint list is vacuously
+// true. The first evaluation error short-circuits and is returned.
+func EvaluateAll(constraints []string, node NodeInfo) (bool, error) {
+	for _, c := range constraints {
+		ok, err := EvaluateConstraint(c, node)
+		if err != nil {
+			return false, fmt.Errorf("constraint %q: %w", c, err)
+		}
+		if !ok {
+			return false, nil
+		}
+	}
+	return true, nil
+}
+
+// ----------------------------------------------------------------------------
+// Value model
+// ----------------------------------------------------------------------------
+
+// celValue is the union of values the evaluator produces. We use the
+// Go interface{} representation so that comparisons can be polymorphic
+// without a tagged-union ceremony; the supported concrete types are
+// bool, int64, and string. Lists are []celValue of the above.
+type celValue = interface{}
+
+// ----------------------------------------------------------------------------
+// Tokenizer
+// ----------------------------------------------------------------------------
+
+type tokKind int
+
+const (
+	tokEOF tokKind = iota
+	tokIdent
+	tokInt
+	tokStr
+	tokOp  // ==, !=, >=, <=, >, <, (, ), .
+	tokIn  // "in"
+	tokAnd // "and"
+	tokOr  // "or"
+	tokNot // "not"
+)
+
+type token struct {
+	kind tokKind
+	text string
+}
+
+type lexer struct {
+	src string
+	pos int
+}
+
+func (l *lexer) next() (token, error) {
+	for l.pos < len(l.src) && unicode.IsSpace(rune(l.src[l.pos])) {
+		l.pos++
+	}
+	if l.pos >= len(l.src) {
+		return token{kind: tokEOF}, nil
+	}
+	c := l.src[l.pos]
+	// string literal
+	if c == '"' {
+		start := l.pos
+		l.pos++
+		for l.pos < len(l.src) && l.src[l.pos] != '"' {
+			l.pos++
+		}
+		if l.pos >= len(l.src) {
+			return token{}, fmt.Errorf("cel: unterminated string at %d", start)
+		}
+		val := l.src[start+1 : l.pos]
+		l.pos++ // consume closing quote
+		return token{kind: tokStr, text: val}, nil
+	}
+	// integer literal
+	if unicode.IsDigit(rune(c)) {
+		start := l.pos
+		for l.pos < len(l.src) && unicode.IsDigit(rune(l.src[l.pos])) {
+			l.pos++
+		}
+		return token{kind: tokInt, text: l.src[start:l.pos]}, nil
+	}
+	// identifier / keyword
+	if isIdentStart(c) {
+		start := l.pos
+		for l.pos < len(l.src) && isIdentPart(l.src[l.pos]) {
+			l.pos++
+		}
+		word := l.src[start:l.pos]
+		switch word {
+		case "in":
+			return token{kind: tokIn, text: word}, nil
+		case "and":
+			return token{kind: tokAnd, text: word}, nil
+		case "or":
+			return token{kind: tokOr, text: word}, nil
+		case "not":
+			return token{kind: tokNot, text: word}, nil
+		default:
+			return token{kind: tokIdent, text: word}, nil
+		}
+	}
+	// operators
+	if strings.ContainsRune("()=!<>.", rune(c)) {
+		// multi-char operators
+		if l.pos+1 < len(l.src) {
+			two := l.src[l.pos : l.pos+2]
+			switch two {
+			case "==", "!=", ">=", "<=":
+				l.pos += 2
+				return token{kind: tokOp, text: two}, nil
+			}
+		}
+		l.pos++
+		return token{kind: tokOp, text: string(c)}, nil
+	}
+	return token{}, fmt.Errorf("cel: unexpected character %q at %d", c, l.pos)
+}
+
+func isIdentStart(c byte) bool {
+	return c == '_' || (c >= 'a' && c <= 'z') || (c >= 'A' && c <= 'Z')
+}
+
+func isIdentPart(c byte) bool {
+	return isIdentStart(c) || (c >= '0' && c <= '9')
+}
+
+// ----------------------------------------------------------------------------
+// Parser (recursive descent, precedence climbing)
+// ----------------------------------------------------------------------------
+
+type parser struct {
+	src  string
+	pos  int
+	tok  token
+	err  error
+	node NodeInfo
+}
+
+func newParser(src string, node NodeInfo) *parser {
+	p := &parser{src: src, node: node}
+	p.advance()
+	return p
+}
+
+func (p *parser) len() int { return len(p.src) }
+
+func (p *parser) advance() {
+	if p.err != nil {
+		return
+	}
+	l := lexer{src: p.src, pos: p.pos}
+	t, err := l.next()
+	if err != nil {
+		p.err = err
+		return
+	}
+	p.pos = l.pos
+	p.tok = t
+}
+
+// Grammar (lowest precedence first):
+//
+//	expr     := orExpr
+//	orExpr   := andExpr ("or" andExpr)*
+//	andExpr  := notExpr ("and" notExpr)*
+//	notExpr  := "not" notExpr | cmpExpr
+//	cmpExpr  := primary (op primary | "in" primary | "not" "in" primary)?
+//	primary  := "(" expr ")"
+//	          | int
+//	          | str
+//	          | "true" | "false"
+//	          | nodeAttr  ("." ident)?   // node.<field>
+//	          | ident                    // bare attribute (e.g. region)
+//	nodeAttr := "node"
+
+func (p *parser) parseExpr() (celValue, error) {
+	if p.err != nil {
+		return nil, p.err
+	}
+	return p.parseOr()
+}
+
+func (p *parser) parseOr() (celValue, error) {
+	left, err := p.parseAnd()
+	if err != nil {
+		return nil, err
+	}
+	for p.tok.kind == tokOr {
+		p.advance()
+		right, err := p.parseAnd()
+		if err != nil {
+			return nil, err
+		}
+		lb, ok := left.(bool)
+		if !ok {
+			return nil, fmt.Errorf("cel: 'or' operand not bool: %T", left)
+		}
+		rb, ok := right.(bool)
+		if !ok {
+			return nil, fmt.Errorf("cel: 'or' operand not bool: %T", right)
+		}
+		left = lb || rb
+	}
+	return left, nil
+}
+
+func (p *parser) parseAnd() (celValue, error) {
+	left, err := p.parseNot()
+	if err != nil {
+		return nil, err
+	}
+	for p.tok.kind == tokAnd {
+		p.advance()
+		right, err := p.parseNot()
+		if err != nil {
+			return nil, err
+		}
+		lb, ok := left.(bool)
+		if !ok {
+			return nil, fmt.Errorf("cel: 'and' operand not bool: %T", left)
+		}
+		rb, ok := right.(bool)
+		if !ok {
+			return nil, fmt.Errorf("cel: 'and' operand not bool: %T", right)
+		}
+		left = lb && rb
+	}
+	return left, nil
+}
+
+func (p *parser) parseNot() (celValue, error) {
+	if p.tok.kind == tokNot {
+		// "not" at the start of a primary is logical negation. "not in"
+		// is handled in parseCmp where it follows a primary.
+		p.advance()
+		v, err := p.parseNot()
+		if err != nil {
+			return nil, err
+		}
+		b, ok := v.(bool)
+		if !ok {
+			return nil, fmt.Errorf("cel: 'not' operand not bool: %T", v)
+		}
+		return !b, nil
+	}
+	return p.parseCmp()
+}
+
+func (p *parser) parseCmp() (celValue, error) {
+	left, err := p.parsePrimary()
+	if err != nil {
+		return nil, err
+	}
+	// "not in"
+	if p.tok.kind == tokNot {
+		p.advance()
+		if p.tok.kind != tokIn {
+			return nil, fmt.Errorf("cel: expected 'in' after 'not', got %q", p.tok.text)
+		}
+		p.advance()
+		right, err := p.parsePrimary()
+		if err != nil {
+			return nil, err
+		}
+		member, err := inMember(left, right)
+		if err != nil {
+			return nil, err
+		}
+		return !member, nil
+	}
+	// "in"
+	if p.tok.kind == tokIn {
+		p.advance()
+		right, err := p.parsePrimary()
+		if err != nil {
+			return nil, err
+		}
+		return inMember(left, right)
+	}
+	// comparison operators
+	if p.tok.kind == tokOp {
+		op := p.tok.text
+		switch op {
+		case "==", "!=", ">=", "<=", ">", "<":
+			p.advance()
+			right, err := p.parsePrimary()
+			if err != nil {
+				return nil, err
+			}
+			return compare(op, left, right)
+		default:
+			return nil, fmt.Errorf("cel: unexpected operator %q", op)
+		}
+	}
+	return left, nil
+}
+
+// inMember reports whether left is a member of right. right must be a
+// list ([]celValue) of comparable values; left may be a string or
+// int64.
+func inMember(left, right celValue) (bool, error) {
+	list, ok := right.([]celValue)
+	if !ok {
+		return false, fmt.Errorf("cel: 'in' rhs not a list: %T", right)
+	}
+	for _, e := range list {
+		if valuesEqual(left, e) {
+			return true, nil
+		}
+	}
+	return false, nil
+}
+
+func valuesEqual(a, b celValue) bool {
+	switch av := a.(type) {
+	case string:
+		bv, ok := b.(string)
+		return ok && av == bv
+	case int64:
+		bv, ok := b.(int64)
+		return ok && av == bv
+	case bool:
+		bv, ok := b.(bool)
+		return ok && av == bv
+	}
+	return false
+}
+
+// compare applies a binary comparison operator to two scalar values.
+// Strings compare lexicographically; ints numerically; bools only via
+// ==/!=.
+func compare(op string, left, right celValue) (bool, error) {
+	switch op {
+	case "==":
+		return valuesEqual(left, right), nil
+	case "!=":
+		return !valuesEqual(left, right), nil
+	}
+	// ordered comparisons require ordered operands
+	ls, lok := left.(string)
+	rs, rok := right.(string)
+	if lok && rok {
+		switch op {
+		case "<":
+			return ls < rs, nil
+		case "<=":
+			return ls <= rs, nil
+		case ">":
+			return ls > rs, nil
+		case ">=":
+			return ls >= rs, nil
+		}
+	}
+	li, lok := left.(int64)
+	ri, rok := right.(int64)
+	if lok && rok {
+		switch op {
+		case "<":
+			return li < ri, nil
+		case "<=":
+			return li <= ri, nil
+		case ">":
+			return li > ri, nil
+		case ">=":
+			return li >= ri, nil
+		}
+	}
+	return false, fmt.Errorf("cel: cannot apply %q to %T and %T", op, left, right)
+}
+
+// parsePrimary parses the smallest standalone unit: parenthesised
+// expressions, literals, and attribute references.
+func (p *parser) parsePrimary() (celValue, error) {
+	switch p.tok.kind {
+	case tokOp:
+		if p.tok.text == "(" {
+			p.advance()
+			v, err := p.parseExpr()
+			if err != nil {
+				return nil, err
+			}
+			if p.tok.kind != tokOp || p.tok.text != ")" {
+				return nil, fmt.Errorf("cel: expected ')' got %q", p.tok.text)
+			}
+			p.advance()
+			return v, nil
+		}
+		return nil, fmt.Errorf("cel: unexpected operator %q", p.tok.text)
+	case tokInt:
+		n, err := strconv.ParseInt(p.tok.text, 10, 64)
+		if err != nil {
+			return nil, fmt.Errorf("cel: bad int %q: %w", p.tok.text, err)
+		}
+		p.advance()
+		return n, nil
+	case tokStr:
+		v := p.tok.text
+		p.advance()
+		return v, nil
+	case tokIdent:
+		return p.parseAttrRef()
+	}
+	return nil, fmt.Errorf("cel: unexpected token %q", p.tok.text)
+}
+
+// parseAttrRef resolves a bare or `node.<field>` attribute reference
+// against the node being evaluated. Bare identifiers (e.g. `region`)
+// resolve against the same attribute map as `node.region`; the PRD
+// examples use both forms interchangeably (see
+// TestParseMarkdown_ConstraintsInlineArray).
+func (p *parser) parseAttrRef() (celValue, error) {
+	name := p.tok.text
+	p.advance()
+	// dotted access: node.<field>
+	if p.tok.kind == tokOp && p.tok.text == "." {
+		if name != "node" {
+			return nil, fmt.Errorf("cel: dotted access on non-node: %q", name)
+		}
+		p.advance()
+		if p.tok.kind != tokIdent {
+			return nil, fmt.Errorf("cel: expected attribute name after '.', got %q", p.tok.text)
+		}
+		field := p.tok.text
+		p.advance()
+		return p.nodeAttr(name + "." + field)
+	}
+	// bare identifier
+	switch name {
+	case "true":
+		return true, nil
+	case "false":
+		return false, nil
+	default:
+		return p.nodeAttr(name)
+	}
+}
+
+// nodeAttr resolves an attribute name to its value on the parser's
+// active node. Mapping (per PRD T2):
+//
+//	node.hostname -> Hostname (string)
+//	node.kind     -> Kind     (string)
+//	node.cpus     -> CPU      (int64)
+//	node.memory   -> Memory   (int64)
+//	node.tags     -> Tags     ([]string -> []celValue)
+//	node.runtimes -> Runtimes ([]string -> []celValue)
+//
+// Bare names (without the `node.` prefix) resolve through the same
+// map, so `region == "us"` and `node.region == "us"` are equivalent
+// when the attribute exists.
+func (p *parser) nodeAttr(name string) (celValue, error) {
+	switch name {
+	case "node.hostname", "hostname":
+		return p.node.Hostname, nil
+	case "node.kind", "kind":
+		return p.node.Kind, nil
+	case "node.cpus", "cpus":
+		return p.node.CPU, nil
+	case "node.memory", "memory":
+		return p.node.Memory, nil
+	case "node.tags", "tags":
+		return toStringValues(p.node.Tags), nil
+	case "node.runtimes", "runtimes":
+		return toStringValues(p.node.Runtimes), nil
+	}
+	return nil, fmt.Errorf("cel: unknown attribute %q", name)
+}
+
+// toStringValues converts a []string to []celValue so the membership
+// operators can compare element-wise.
+func toStringValues(in []string) []celValue {
+	out := make([]celValue, len(in))
+	for i, s := range in {
+		out[i] = s
+	}
+	return out
+}

internal/scheduler/cel_test.go

@@ -0,0 +1,210 @@
+package scheduler
+
+import "testing"
+
+func TestEvaluateConstraint_Equality(t *testing.T) {
+	node := NodeInfo{Hostname: "h-1", Kind: "linux", CPU: 4, Memory: 4096, Tags: []string{"web"}, Runtimes: []string{"process"}}
+	cases := []struct {
+		name string
+		expr string
+		want bool
+	}{
+		{"hostname eq", `node.hostname == "h-1"`, true},
+		{"hostname ne", `node.hostname == "h-2"`, false},
+		{"kind eq", `node.kind == "linux"`, true},
+		{"kind ne", `node.kind == "proxmox"`, false},
+		{"cpus eq", `node.cpus == 4`, true},
+		{"memory eq", `node.memory == 4096`, true},
+	}
+	for _, c := range cases {
+		got, err := EvaluateConstraint(c.expr, node)
+		if err != nil {
+			t.Errorf("%s: %v", c.name, err)
+			continue
+		}
+		if got != c.want {
+			t.Errorf("%s: got %v, want %v", c.name, got, c.want)
+		}
+	}
+}
+
+func TestEvaluateConstraint_Comparison(t *testing.T) {
+	node := NodeInfo{Hostname: "h", Kind: "linux", CPU: 4, Memory: 4096}
+	cases := []struct {
+		expr string
+		want bool
+	}{
+		{"node.cpus >= 2", true},
+		{"node.cpus >= 4", true},
+		{"node.cpus > 4", false},
+		{"node.cpus > 2", true},
+		{"node.cpus <= 4", true},
+		{"node.cpus < 2", false},
+		{"node.cpus != 8", true},
+		{"node.cpus == 8", false},
+		{"node.memory >= 2048", true},
+		{"node.memory < 1024", false},
+	}
+	for _, c := range cases {
+		got, err := EvaluateConstraint(c.expr, node)
+		if err != nil {
+			t.Errorf("%q: %v", c.expr, err)
+			continue
+		}
+		if got != c.want {
+			t.Errorf("%q: got %v, want %v", c.expr, got, c.want)
+		}
+	}
+}
+
+func TestEvaluateConstraint_Membership(t *testing.T) {
+	node := NodeInfo{Tags: []string{"web", "log-shipper"}, Runtimes: []string{"process", "wasmtime"}}
+	cases := []struct {
+		expr string
+		want bool
+	}{
+		{`"web" in node.tags`, true},
+		{`"missing" in node.tags`, false},
+		{`"process" in node.runtimes`, true},
+		{`"podman" in node.runtimes`, false},
+		{`"log-shipper" not in node.tags`, false},
+		{`"missing" not in node.tags`, true},
+	}
+	for _, c := range cases {
+		got, err := EvaluateConstraint(c.expr, node)
+		if err != nil {
+			t.Errorf("%q: %v", c.expr, err)
+			continue
+		}
+		if got != c.want {
+			t.Errorf("%q: got %v, want %v", c.expr, got, c.want)
+		}
+	}
+}
+
+func TestEvaluateConstraint_BooleanComposition(t *testing.T) {
+	node := NodeInfo{Kind: "linux", CPU: 4, Tags: []string{"web"}}
+	cases := []struct {
+		expr string
+		want bool
+	}{
+		{`node.kind == "linux" and node.cpus >= 2`, true},
+		{`node.kind == "proxmox" and node.cpus >= 2`, false},
+		{`node.kind == "linux" or node.kind == "proxmox"`, true},
+		{`node.kind == "proxmox" or node.kind == "linux"`, true},
+		{`not node.kind == "proxmox"`, true},
+		{`not node.kind == "linux"`, false},
+		{`(node.kind == "linux") and (node.cpus >= 2)`, true},
+		{`node.cpus >= 2 and not "blocked" in node.tags`, true},
+		{`node.kind == "linux" and node.cpus >= 2 and "web" in node.tags`, true},
+		{`node.kind == "linux" or node.kind == "proxmox" or node.cpus > 100`, true},
+	}
+	for _, c := range cases {
+		got, err := EvaluateConstraint(c.expr, node)
+		if err != nil {
+			t.Errorf("%q: %v", c.expr, err)
+			continue
+		}
+		if got != c.want {
+			t.Errorf("%q: got %v, want %v", c.expr, got, c.want)
+		}
+	}
+}
+
+func TestEvaluateConstraint_BareIdentifiers(t *testing.T) {
+	// Bare identifiers resolve through the same attribute map as
+	// node.<field> (per PRD: constraints may use either form).
+	node := NodeInfo{Kind: "linux", CPU: 4}
+	got, err := EvaluateConstraint(`kind == "linux"`, node)
+	if err != nil {
+		t.Fatalf("bare kind: %v", err)
+	}
+	if !got {
+		t.Error("bare kind == linux: got false, want true")
+	}
+}
+
+func TestEvaluateConstraint_TrueFalseLiterals(t *testing.T) {
+	node := NodeInfo{}
+	cases := []struct {
+		expr string
+		want bool
+	}{
+		{"true", true},
+		{"false", false},
+		{"not false", true},
+		{"not true", false},
+		{"true and true", true},
+		{"true and false", false},
+		{"false or true", true},
+	}
+	for _, c := range cases {
+		got, err := EvaluateConstraint(c.expr, node)
+		if err != nil {
+			t.Errorf("%q: %v", c.expr, err)
+			continue
+		}
+		if got != c.want {
+			t.Errorf("%q: got %v, want %v", c.expr, got, c.want)
+		}
+	}
+}
+
+func TestEvaluateConstraint_Errors(t *testing.T) {
+	node := NodeInfo{Kind: "linux"}
+	cases := []struct {
+		name string
+		expr string
+	}{
+		{"empty", ""},
+		{"unterminated string", `node.kind == "linux`},
+		{"unknown attribute", `node.bogus == 1`},
+		{"unknown bare attr", `bogus == 1`},
+		{"dotted on non-node", `host.kind == "linux"`},
+		{"bad operator", `node.cpus + 2`},
+		{"trailing input", `node.kind == "linux" garbage`},
+		{"unbalanced paren", `(node.kind == "linux"`},
+		{"missing rhs", `node.cpus >=`},
+		{"not without in", `"x" not node.tags`},
+		{"ordered compare on bool", `true < false`},
+		{"ordered compare on mismatched types", `node.kind > 2`},
+		{"in on non-list", `"x" in node.kind`},
+	}
+	for _, c := range cases {
+		_, err := EvaluateConstraint(c.expr, node)
+		if err == nil {
+			t.Errorf("%s: expected error for %q, got nil", c.name, c.expr)
+		}
+	}
+}
+
+func TestEvaluateAll(t *testing.T) {
+	node := NodeInfo{Kind: "linux", CPU: 4, Tags: []string{"web"}}
+	cases := []struct {
+		name        string
+		constraints []string
+		want        bool
+	}{
+		{"empty", nil, true},
+		{"all pass", []string{`node.kind == "linux"`, "node.cpus >= 2"}, true},
+		{"one fails", []string{`node.kind == "linux"`, "node.cpus >= 8"}, false},
+		{"all fail", []string{`node.kind == "proxmox"`, "node.cpus >= 8"}, false},
+	}
+	for _, c := range cases {
+		got, err := EvaluateAll(c.constraints, node)
+		if err != nil {
+			t.Errorf("%s: %v", c.name, err)
+			continue
+		}
+		if got != c.want {
+			t.Errorf("%s: got %v, want %v", c.name, got, c.want)
+		}
+	}
+}
+
+func TestEvaluateAll_PropagatesError(t *testing.T) {
+	node := NodeInfo{}
+	if _, err := EvaluateAll([]string{"bogus == 1"}, node); err == nil {
+		t.Error("EvaluateAll: expected error for malformed constraint")
+	}
+}

internal/scheduler/scheduler.go

@@ -0,0 +1,472 @@
+// Package scheduler implements the v0.9 CLI-side scheduler (REQ-083,
+// P05). Unlike the v0.8 daemon-side best-fit scheduler
+// (internal/engine/scheduler.go), this scheduler runs entirely in the
+// `orca` CLI process (R-001) and is pure: it takes a list of candidate
+// nodes plus a workload request and returns placement decisions
+// without performing any I/O.
+//
+// The scheduler is runtime-aware: a workload that declares
+// `runtime.one_of: wasm` is only placed on nodes that expose
+// `wasmtime` in their Runtimes list; a `pve-vm` workload is only
+// placed on `proxmox` nodes. It is also constraint- and
+// affinity-aware via the CEL-subset evaluator in cel.go.
+//
+// Workload kinds are handled differently per the PRD:
+//
+//   - Job: one-shot, returns exactly one placement (best-fit
+//     bin-packing).
+//   - Service: count replicas spread across distinct nodes
+//     (anti-affinity by default); if fewer distinct nodes than count,
+//     colocation is permitted but distinct nodes are preferred.
+//   - DaemonSet: one placement per node that fits the constraints.
+package scheduler
+
+import (
+	"fmt"
+	"sort"
+
+	"git.cloudinit.dev/coreci/orca/internal/jobspec"
+)
+
+// NodeInfo is the scheduler's projection of a peer node: total and
+// free capacity, the runtimes the node advertises, its tags, and its
+// kind (linux/proxmox). The CLI populates this from the
+// cluster/peers/ inventory plus the per-node capacity reports
+// collected over SSH; the scheduler itself never reads either.
+type NodeInfo struct {
+	Hostname string
+	Runtimes []string
+	Tags     []string
+	CPU      int64
+	Memory   int64
+	FreeCPU  int64
+	FreeMem  int64
+	Kind     string
+}
+
+// WorkloadRequest bundles a parsed WorkloadSpec with the namespace
+// the workload is being scheduled into. The namespace is carried
+// through to placement so the resulting AllocID can be namespaced,
+// but the scheduler itself does not inspect it for fitting decisions.
+type WorkloadRequest struct {
+	Spec      *jobspec.WorkloadSpec
+	Namespace string
+}
+
+// Placement is a single scheduling decision: which node, which
+// allocation id, and the bin-packing score that won the node the
+// placement. AllocID is `ns/spec.Name-<idx>` so a multi-replica
+// Service produces distinct ids per replica.
+type Placement struct {
+	Node    string
+	AllocID string
+	Score   int64
+}
+
+// Schedule is the main entry point. For Job (kind=Job) it returns one
+// placement on the best-fit node. For Service it returns `Count`
+// placements spread across distinct nodes where possible (anti-
+// affinity), permitting colocation when Count > nodes. For DaemonSet
+// it returns one placement per node that fits. Any kind-agnostic
+// validation error (no spec, unknown kind, no fitting node) is
+// returned as an error rather than an empty slice so callers can
+// distinguish "nothing fits" from "scheduled zero replicas".
+func Schedule(nodes []NodeInfo, req WorkloadRequest) ([]Placement, error) {
+	if req.Spec == nil {
+		return nil, fmt.Errorf("scheduler: nil WorkloadSpec")
+	}
+	if len(nodes) == 0 {
+		return nil, fmt.Errorf("scheduler: no candidate nodes")
+	}
+	switch req.Spec.Kind {
+	case "Job":
+		return scheduleJob(nodes, req)
+	case "Service":
+		return scheduleService(nodes, req)
+	case "DaemonSet":
+		return scheduleDaemonSet(nodes, req)
+	default:
+		return nil, fmt.Errorf("scheduler: unknown kind %q", req.Spec.Kind)
+	}
+}
+
+// Score evaluates a single node against a workload. fits is true iff
+// the node (a) advertises a runtime compatible with the workload's
+// `runtime.one_of`, (b) satisfies every CEL constraint in
+// `spec.Constraints`, and (c) has enough free CPU+memory for the
+// workload's requested resources. When fits is true, score is the
+// bin-packing score (more free capacity = higher score, so the node
+// most likely to absorb the workload without starving its
+// neighbours wins). When fits is false, score is 0.
+func Score(node NodeInfo, req WorkloadRequest) (score int64, fits bool) {
+	// (a) runtime compatibility. A workload with no Runtime block or
+	// an empty OneOf is treated as runtime-agnostic (always fits on
+	// the runtime axis); this matches the v0.8 behaviour where a
+	// missing runtime meant "process".
+	runtimeOK := true
+	if req.Spec != nil && req.Spec.Runtime != nil && req.Spec.Runtime.OneOf != "" {
+		runtimeOK = hasRuntime(node, req.Spec.Runtime.OneOf)
+	}
+	if !runtimeOK {
+		return 0, false
+	}
+	// (b) constraints. Evaluation errors are treated as non-fit so a
+	// malformed constraint does not crash Schedule; the caller still
+	// sees the node filtered out.
+	if req.Spec != nil {
+		ok, err := EvaluateAll(req.Spec.Constraints, node)
+		if err != nil || !ok {
+			return 0, false
+		}
+	}
+	// (c) capacity. A workload with no Resources block is treated as
+	// zero-sized for fitting purposes (it always fits the capacity
+	// axis); real workloads declare cpu/memory.
+	needCPU, needMem := workloadResources(req)
+	if node.FreeCPU < needCPU || node.FreeMem < needMem {
+		return 0, false
+	}
+	// bin-packing score: most free capacity wins. CPU is weighted
+	// 1000x memory so a 1-core difference outweighs a 1-MiB
+	// difference, mirroring the v0.8 Score weighting that biased
+	// toward CPU (the more common binding constraint).
+	score = (node.FreeCPU-needCPU)*1000 + (node.FreeMem - needMem)
+	if score < 0 {
+		score = 0
+	}
+	return score, true
+}
+
+// hasRuntime reports whether node advertises the requested runtime.
+// The match is case-insensitive and tolerant of aliases: `wasm` and
+// `wasmtime` are treated as the same runtime, and `pve-vm`/`pve-ct`
+// only match nodes whose Kind is "proxmox".
+func hasRuntime(node NodeInfo, oneOf string) bool {
+	want := normalizeRuntime(oneOf)
+	// pve-* runtimes require a proxmox-kind node regardless of the
+	// node's Runtimes list (a proxmox node doesn't list "pve-vm" in
+	// Runtimes; it IS the runtime).
+	switch want {
+	case "pve-vm", "pve-ct", "proxmox":
+		return normalizeKind(node.Kind) == "proxmox"
+	}
+	for _, r := range node.Runtimes {
+		if normalizeRuntime(r) == want {
+			return true
+		}
+		// alias: wasmtime nodes advertise "wasmtime"; workloads ask
+		// for "wasm".
+		if want == "wasm" && normalizeRuntime(r) == "wasmtime" {
+			return true
+		}
+	}
+	return false
+}
+
+// normalizeRuntime lowercases and trims a runtime name for matching.
+func normalizeRuntime(s string) string {
+	s = toLowerASCII(s)
+	switch s {
+	case "wasmtime":
+		return "wasm"
+	}
+	return s
+}
+
+// normalizeKind lowercases and trims a node Kind for matching.
+func normalizeKind(s string) string { return toLowerASCII(s) }
+
+// toLowerASCII lowercases ASCII letters without bringing in strings
+// (avoid an alloc-heavy stdlib call in the hot path).
+func toLowerASCII(s string) string {
+	b := []byte(s)
+	for i, c := range b {
+		if c >= 'A' && c <= 'Z' {
+			b[i] = c + 32
+		}
+	}
+	return string(b)
+}
+
+// workloadResources returns the (cpu, memory) the workload requests,
+// read from the WorkloadSpec's Resources block if present. The
+// v0.9-P05 WorkloadSpec does not yet carry a Resources field (it
+// lands in P0c, REQ-074); until then this returns (0, 0) so the
+// capacity check is a no-op and runtime/constraints do the real
+// filtering. The signature is here so the scheduler logic does not
+// need to change when Resources lands.
+func workloadResources(req WorkloadRequest) (int64, int64) {
+	_ = req
+	return 0, 0
+}
+
+// ----------------------------------------------------------------------------
+// Kind-specific scheduling
+// ----------------------------------------------------------------------------
+
+// scheduleJob places a single Job on the best-fit node.
+func scheduleJob(nodes []NodeInfo, req WorkloadRequest) ([]Placement, error) {
+	type cand struct {
+		node  NodeInfo
+		score int64
+	}
+	var cands []cand
+	for _, n := range nodes {
+		s, ok := Score(n, req)
+		if !ok {
+			continue
+		}
+		cands = append(cands, cand{node: n, score: s})
+	}
+	// CEL-based affinity rules apply to single-shot Jobs too: a Job
+	// with `affinity: [{target: "\"ssd\" in node.tags", weight: 100}]`
+	// should land on the tagged node even without prior placements.
+	// Name-based affinity (no prior placements to check) contributes
+	// zero for a standalone Job, so it is harmless to call here.
+	for i := range cands {
+		cands[i].score += affinityScore(cands[i].node, req, nil)
+	}
+	if len(cands) == 0 {
+		return nil, fmt.Errorf("scheduler: no node fits workload %q", req.Spec.Name)
+	}
+	sort.SliceStable(cands, func(i, j int) bool {
+		if cands[i].score != cands[j].score {
+			return cands[i].score > cands[j].score
+		}
+		return cands[i].node.Hostname < cands[j].node.Hostname
+	})
+	w := cands[0]
+	return []Placement{{
+		Node:    w.node.Hostname,
+		AllocID: allocID(req, 0),
+		Score:   w.score,
+	}}, nil
+}
+
+// scheduleService places `Count` replicas with implicit anti-affinity:
+// prefer distinct nodes, but permit colocation when Count exceeds the
+// number of fitting nodes. Each replica gets a distinct AllocID.
+func scheduleService(nodes []NodeInfo, req WorkloadRequest) ([]Placement, error) {
+	count := req.Spec.Count
+	if count <= 0 {
+		count = 1
+	}
+	// Pre-filter fitting nodes once; the loop below re-scores them
+	// after each placement so the capacity accounting reflects the
+	// replicas already placed.
+	fitting := filterFitting(nodes, req)
+	if len(fitting) == 0 {
+		return nil, fmt.Errorf("scheduler: no node fits service %q", req.Spec.Name)
+	}
+	var placements []Placement
+	placed := map[string]int{} // hostname -> count placed there
+	// First pass: spread across distinct nodes.
+	for i := 0; i < count; i++ {
+		best, score, ok := pickServiceNode(fitting, req, placements, placed)
+		if !ok {
+			break
+		}
+		placements = append(placements, Placement{
+			Node:    best.Hostname,
+			AllocID: allocID(req, i),
+			Score:   score,
+		})
+		placed[best.Hostname]++
+		// Reflect the consumed capacity in the candidate snapshot so
+		// subsequent picks see updated free capacity.
+		needCPU, needMem := workloadResources(req)
+		for j := range fitting {
+			if fitting[j].Hostname == best.Hostname {
+				fitting[j].FreeCPU -= needCPU
+				fitting[j].FreeMem -= needMem
+			}
+		}
+	}
+	if len(placements) < count {
+		return nil, fmt.Errorf("scheduler: only placed %d/%d replicas for service %q",
+			len(placements), count, req.Spec.Name)
+	}
+	return placements, nil
+}
+
+// pickServiceNode selects the best node for the next replica. The
+// selection prefers nodes with zero prior placements of this service
+// (anti-affinity) and applies affinity scoring on top of the
+// bin-packing score.
+func pickServiceNode(fitting []NodeInfo, req WorkloadRequest, placements []Placement, placed map[string]int) (NodeInfo, int64, bool) {
+	type scored struct {
+		node  NodeInfo
+		score int64
+	}
+	var cands []scored
+	for _, n := range fitting {
+		s, ok := Score(n, req)
+		if !ok {
+			continue
+		}
+		// Implicit anti-affinity: a node with N prior replicas of this
+		// service incurs a penalty of N * (1 << 62) so distinct nodes
+		// are preferred, but colocation is permitted (with a
+		// per-replica penalty) when no distinct node remains. This
+		// produces a balanced spread (e.g. 5 replicas on 3 nodes →
+		// 2/2/1) rather than stacking everything on the first node.
+		if placed[n.Hostname] > 0 {
+			s -= int64(placed[n.Hostname]) * (1 << 62)
+		}
+		// Affinity rules from the spec add/subtract their weight.
+		s += affinityScore(n, req, placements)
+		cands = append(cands, scored{node: n, score: s})
+	}
+	if len(cands) == 0 {
+		return NodeInfo{}, 0, false
+	}
+	sort.SliceStable(cands, func(i, j int) bool {
+		if cands[i].score != cands[j].score {
+			return cands[i].score > cands[j].score
+		}
+		return cands[i].node.Hostname < cands[j].node.Hostname
+	})
+	w := cands[0]
+	return w.node, w.score, true
+}
+
+// scheduleDaemonSet places one replica per node that fits the
+// constraints. The PRD's DaemonSet placement mode (every-node /
+// matching / mandatory) lives on the WorkloadSpec.Schedule block; the
+// scheduler honours it indirectly by filtering on Constraints: a
+// `matching` DaemonSet carries constraints that select the matching
+// nodes, an `every-node` DaemonSet carries none, and a `mandatory`
+// one is enforced elsewhere (the scheduler still just returns
+// placements for every fitting node).
+func scheduleDaemonSet(nodes []NodeInfo, req WorkloadRequest) ([]Placement, error) {
+	var placements []Placement
+	for _, n := range nodes {
+		s, ok := Score(n, req)
+		if !ok {
+			continue
+		}
+		placements = append(placements, Placement{
+			Node:    n.Hostname,
+			AllocID: allocID(req, len(placements)),
+			Score:   s,
+		})
+	}
+	if len(placements) == 0 {
+		return nil, fmt.Errorf("scheduler: no node fits daemonset %q", req.Spec.Name)
+	}
+	return placements, nil
+}
+
+// ----------------------------------------------------------------------------
+// Affinity scoring
+// ----------------------------------------------------------------------------
+
+// affinityScore returns the weighted affinity contribution for a
+// node given the placements already made. For each AffinityRule the
+// Target is a CEL expression; if it evaluates true against the node,
+// the rule's Weight is added (positive = co-locate, negative =
+// anti-affinity). An affinity target that fails to evaluate is
+// ignored rather than failing the schedule: operators use affinity as
+// a hint, not a hard gate.
+//
+// The PRD also mentions affinity rules like `{target: "redis", weight:
+// 50}` where Target is a workload *name* rather than a CEL expression.
+// We support both: if Target parses as a CEL expression it is
+// evaluated against the node; otherwise it is treated as a workload
+// name and we check whether any already-placed alloc for that name
+// exists on the node. The placement-already-here check is done by the
+// caller via placements; this function checks the node's own
+// attributes only.
+func affinityScore(node NodeInfo, req WorkloadRequest, placements []Placement) int64 {
+	if req.Spec == nil {
+		return 0
+	}
+	var total int64
+	for _, rule := range req.Spec.Affinity {
+		// Try CEL evaluation first; if the target is a bare workload
+		// name (no operator) the CEL parser will fail and we fall
+		// back to name-based placement counting.
+		ok, err := EvaluateConstraint(rule.Target, node)
+		if err == nil {
+			if ok {
+				total += int64(rule.Weight)
+			}
+			continue
+		}
+		// Fallback: target is a workload name; count existing
+		// placements for that workload on this node and apply the
+		// weight once per co-located replica.
+		for _, p := range placements {
+			if p.Node == node.Hostname && isAllocFor(p.AllocID, rule.Target) {
+				total += int64(rule.Weight)
+			}
+		}
+	}
+	return total
+}
+
+// isAllocFor reports whether an AllocID encodes a placement for the
+// named workload. AllocIDs are `ns/name-idx`, so we look for the
+// workload name as the segment after the first slash and before the
+// trailing `-idx`.
+func isAllocFor(allocID, workloadName string) bool {
+	// strip namespace prefix
+	rest := allocID
+	if i := indexByte(rest, '/'); i >= 0 {
+		rest = rest[i+1:]
+	}
+	// strip trailing -idx
+	if i := lastIndexByte(rest, '-'); i >= 0 {
+		rest = rest[:i]
+	}
+	return rest == workloadName
+}
+
+// indexByte returns the index of the first occurrence of b in s, or
+// -1. Avoids importing strings just for one helper.
+func indexByte(s string, b byte) int {
+	for i := 0; i < len(s); i++ {
+		if s[i] == b {
+			return i
+		}
+	}
+	return -1
+}
+
+// lastIndexByte returns the index of the last occurrence of b in s, or
+// -1.
+func lastIndexByte(s string, b byte) int {
+	for i := len(s) - 1; i >= 0; i-- {
+		if s[i] == b {
+			return i
+		}
+	}
+	return -1
+}
+
+// ----------------------------------------------------------------------------
+// Helpers
+// ----------------------------------------------------------------------------
+
+// filterFitting returns a copy of the nodes that pass Score for the
+// request, preserving order. Capacity is not yet decremented; the
+// caller adjusts FreeCPU/FreeMem as it places replicas.
+func filterFitting(nodes []NodeInfo, req WorkloadRequest) []NodeInfo {
+	var out []NodeInfo
+	for _, n := range nodes {
+		if _, ok := Score(n, req); ok {
+			out = append(out, n)
+		}
+	}
+	return out
+}
+
+// allocID renders a stable, namespaced allocation id for a placement.
+// Format: `ns/spec.Name-<idx>`.
+func allocID(req WorkloadRequest, idx int) string {
+	ns := req.Namespace
+	if ns == "" {
+		ns = "default"
+	}
+	return fmt.Sprintf("%s/%s-%d", ns, req.Spec.Name, idx)
+}

internal/scheduler/scheduler_test.go

@@ -0,0 +1,451 @@
+package scheduler
+
+import (
+	"strings"
+	"testing"
+
+	"git.cloudinit.dev/coreci/orca/internal/jobspec"
+)
+
+// threeLinuxNodes returns a small cluster of three Linux nodes with
+// distinct free capacities so best-fit ordering is unambiguous.
+func threeLinuxNodes() []NodeInfo {
+	return []NodeInfo{
+		{Hostname: "node-a", Runtimes: []string{"process"}, Tags: nil, CPU: 4, Memory: 4096, FreeCPU: 4, FreeMem: 4096, Kind: "linux"},
+		{Hostname: "node-b", Runtimes: []string{"process"}, Tags: nil, CPU: 8, Memory: 8192, FreeCPU: 8, FreeMem: 8192, Kind: "linux"},
+		{Hostname: "node-c", Runtimes: []string{"process"}, Tags: nil, CPU: 2, Memory: 2048, FreeCPU: 2, FreeMem: 2048, Kind: "linux"},
+	}
+}
+
+func jobSpec(name, oneOf string, constraints []string) *jobspec.WorkloadSpec {
+	return &jobspec.WorkloadSpec{
+		Kind:        "Job",
+		Name:        name,
+		Count:       1,
+		Runtime:     &jobspec.RuntimeBlock{OneOf: oneOf},
+		Constraints: constraints,
+	}
+}
+
+func serviceSpec(name, oneOf string, count int, constraints []string) *jobspec.WorkloadSpec {
+	return &jobspec.WorkloadSpec{
+		Kind:        "Service",
+		Name:        name,
+		Count:       count,
+		Runtime:     &jobspec.RuntimeBlock{OneOf: oneOf},
+		Constraints: constraints,
+	}
+}
+
+func daemonSetSpec(name, oneOf string, constraints []string) *jobspec.WorkloadSpec {
+	return &jobspec.WorkloadSpec{
+		Kind:        "DaemonSet",
+		Name:        name,
+		Count:       1,
+		Runtime:     &jobspec.RuntimeBlock{OneOf: oneOf},
+		Constraints: constraints,
+	}
+}
+
+// ---------------------------------------------------------------------------
+// Job
+// ---------------------------------------------------------------------------
+
+func TestScheduleJob_BestFit(t *testing.T) {
+	nodes := threeLinuxNodes()
+	req := WorkloadRequest{Spec: jobSpec("batch", "process", nil), Namespace: "ns"}
+	got, err := Schedule(nodes, req)
+	if err != nil {
+		t.Fatalf("Schedule: %v", err)
+	}
+	if len(got) != 1 {
+		t.Fatalf("placements = %d, want 1", len(got))
+	}
+	if got[0].Node != "node-b" {
+		t.Errorf("Node = %q, want node-b (most free capacity)", got[0].Node)
+	}
+	if !strings.HasPrefix(got[0].AllocID, "ns/batch-") {
+		t.Errorf("AllocID = %q, want ns/batch-*", got[0].AllocID)
+	}
+	if got[0].Score <= 0 {
+		t.Errorf("Score = %d, want > 0", got[0].Score)
+	}
+}
+
+func TestScheduleJob_NoFittingNode(t *testing.T) {
+	nodes := threeLinuxNodes()
+	// wasm runtime not advertised by any node.
+	req := WorkloadRequest{Spec: jobSpec("wasmjob", "wasm", nil), Namespace: "ns"}
+	if _, err := Schedule(nodes, req); err == nil {
+		t.Fatal("Schedule: expected error for no-fitting node, got nil")
+	}
+}
+
+// ---------------------------------------------------------------------------
+// Service
+// ---------------------------------------------------------------------------
+
+func TestScheduleService_SpreadAcrossNodes(t *testing.T) {
+	nodes := threeLinuxNodes()
+	req := WorkloadRequest{Spec: serviceSpec("web", "process", 3, nil), Namespace: "ns"}
+	got, err := Schedule(nodes, req)
+	if err != nil {
+		t.Fatalf("Schedule: %v", err)
+	}
+	if len(got) != 3 {
+		t.Fatalf("placements = %d, want 3", len(got))
+	}
+	seen := map[string]int{}
+	for _, p := range got {
+		seen[p.Node]++
+	}
+	if len(seen) != 3 {
+		t.Errorf("anti-affinity spread: distinct nodes = %d, want 3; %v", len(seen), seen)
+	}
+}
+
+func TestScheduleService_ColocationWhenFewerNodes(t *testing.T) {
+	nodes := threeLinuxNodes()
+	req := WorkloadRequest{Spec: serviceSpec("web", "process", 5, nil), Namespace: "ns"}
+	got, err := Schedule(nodes, req)
+	if err != nil {
+		t.Fatalf("Schedule: %v", err)
+	}
+	if len(got) != 5 {
+		t.Fatalf("placements = %d, want 5", len(got))
+	}
+	seen := map[string]int{}
+	for _, p := range got {
+		seen[p.Node]++
+	}
+	if len(seen) != 3 {
+		t.Errorf("colocation: distinct nodes = %d, want 3 (all used)", len(seen))
+	}
+	// No node should host more than 2 (3 nodes, 5 replicas: 2+2+1).
+	for n, c := range seen {
+		if c > 2 {
+			t.Errorf("node %s has %d replicas, want <= 2", n, c)
+		}
+	}
+}
+
+func TestScheduleService_NoFittingNode(t *testing.T) {
+	nodes := threeLinuxNodes()
+	req := WorkloadRequest{Spec: serviceSpec("wasm-svc", "wasm", 3, nil), Namespace: "ns"}
+	if _, err := Schedule(nodes, req); err == nil {
+		t.Fatal("Schedule: expected error for service with no fitting node")
+	}
+}
+
+// ---------------------------------------------------------------------------
+// DaemonSet
+// ---------------------------------------------------------------------------
+
+func TestScheduleDaemonSet_AllMatching(t *testing.T) {
+	nodes := threeLinuxNodes()
+	req := WorkloadRequest{Spec: daemonSetSpec("logrotate", "process", nil), Namespace: "ns"}
+	got, err := Schedule(nodes, req)
+	if err != nil {
+		t.Fatalf("Schedule: %v", err)
+	}
+	if len(got) != 3 {
+		t.Errorf("placements = %d, want 3 (one per node)", len(got))
+	}
+	seen := map[string]bool{}
+	for _, p := range got {
+		seen[p.Node] = true
+	}
+	if len(seen) != 3 {
+		t.Errorf("DaemonSet distinct nodes = %d, want 3", len(seen))
+	}
+}
+
+func TestScheduleDaemonSet_SomeExcludedByConstraint(t *testing.T) {
+	nodes := threeLinuxNodes()
+	// Only nodes with cpus >= 4 qualify: node-a (4) and node-b (8).
+	req := WorkloadRequest{Spec: daemonSetSpec("heavy", "process", []string{"node.cpus >= 4"}), Namespace: "ns"}
+	got, err := Schedule(nodes, req)
+	if err != nil {
+		t.Fatalf("Schedule: %v", err)
+	}
+	if len(got) != 2 {
+		t.Errorf("placements = %d, want 2 (cpus>=4)", len(got))
+	}
+}
+
+// ---------------------------------------------------------------------------
+// Runtime compatibility
+// ---------------------------------------------------------------------------
+
+func TestSchedule_RuntimeCompatibilityWasm(t *testing.T) {
+	nodes := []NodeInfo{
+		{Hostname: "no-wasm", Runtimes: []string{"process"}, Kind: "linux", CPU: 8, Memory: 8192, FreeCPU: 8, FreeMem: 8192},
+		{Hostname: "has-wasm", Runtimes: []string{"process", "wasmtime"}, Kind: "linux", CPU: 4, Memory: 4096, FreeCPU: 4, FreeMem: 4096},
+	}
+	// Even though no-wasm has more free capacity, the wasm workload
+	// must land on has-wasm.
+	req := WorkloadRequest{Spec: jobSpec("wasmjob", "wasm", nil), Namespace: "ns"}
+	got, err := Schedule(nodes, req)
+	if err != nil {
+		t.Fatalf("Schedule: %v", err)
+	}
+	if got[0].Node != "has-wasm" {
+		t.Errorf("Node = %q, want has-wasm (runtime compatibility)", got[0].Node)
+	}
+}
+
+func TestSchedule_RuntimeCompatibilityPveVM(t *testing.T) {
+	nodes := []NodeInfo{
+		{Hostname: "linux-1", Runtimes: []string{"process"}, Kind: "linux", CPU: 8, Memory: 8192, FreeCPU: 8, FreeMem: 8192},
+		{Hostname: "pve-1", Runtimes: []string{"process"}, Kind: "proxmox", CPU: 8, Memory: 8192, FreeCPU: 8, FreeMem: 8192},
+	}
+	req := WorkloadRequest{Spec: jobSpec("vmjob", "pve-vm", nil), Namespace: "ns"}
+	got, err := Schedule(nodes, req)
+	if err != nil {
+		t.Fatalf("Schedule: %v", err)
+	}
+	if got[0].Node != "pve-1" {
+		t.Errorf("Node = %q, want pve-1 (pve-vm requires proxmox kind)", got[0].Node)
+	}
+}
+
+// ---------------------------------------------------------------------------
+// Constraints
+// ---------------------------------------------------------------------------
+
+func TestSchedule_ConstraintKindExcludesProxmox(t *testing.T) {
+	nodes := []NodeInfo{
+		{Hostname: "linux-1", Runtimes: []string{"process"}, Kind: "linux", CPU: 8, Memory: 8192, FreeCPU: 8, FreeMem: 8192},
+		{Hostname: "pve-1", Runtimes: []string{"process"}, Kind: "proxmox", CPU: 8, Memory: 8192, FreeCPU: 8, FreeMem: 8192},
+	}
+	req := WorkloadRequest{Spec: jobSpec("linuxonly", "process", []string{`node.kind == "linux"`}), Namespace: "ns"}
+	got, err := Schedule(nodes, req)
+	if err != nil {
+		t.Fatalf("Schedule: %v", err)
+	}
+	if got[0].Node != "linux-1" {
+		t.Errorf("Node = %q, want linux-1 (kind==linux)", got[0].Node)
+	}
+}
+
+func TestSchedule_ConstraintCPUsExcludesSmall(t *testing.T) {
+	nodes := threeLinuxNodes() // node-c has cpus=2
+	req := WorkloadRequest{Spec: jobSpec("big", "process", []string{"node.cpus >= 4"}), Namespace: "ns"}
+	got, err := Schedule(nodes, req)
+	if err != nil {
+		t.Fatalf("Schedule: %v", err)
+	}
+	if got[0].Node == "node-c" {
+		t.Errorf("Node = node-c, want node-a or node-b (cpus>=4)")
+	}
+}
+
+func TestSchedule_ConstraintNotInTags(t *testing.T) {
+	nodes := []NodeInfo{
+		{Hostname: "tagged", Runtimes: []string{"process"}, Tags: []string{"log-shipper"}, Kind: "linux", CPU: 8, Memory: 8192, FreeCPU: 8, FreeMem: 8192},
+		{Hostname: "clean", Runtimes: []string{"process"}, Tags: nil, Kind: "linux", CPU: 4, Memory: 4096, FreeCPU: 4, FreeMem: 4096},
+	}
+	req := WorkloadRequest{Spec: jobSpec("worker", "process", []string{`"log-shipper" not in node.tags`}), Namespace: "ns"}
+	got, err := Schedule(nodes, req)
+	if err != nil {
+		t.Fatalf("Schedule: %v", err)
+	}
+	if got[0].Node != "clean" {
+		t.Errorf("Node = %q, want clean (log-shipper not in tags)", got[0].Node)
+	}
+}
+
+// ---------------------------------------------------------------------------
+// Affinity
+// ---------------------------------------------------------------------------
+
+func TestSchedule_AffinityPrefersColocatedNode(t *testing.T) {
+	// Place a redis service first, then a worker with affinity for
+	// redis; the worker should prefer the node where redis already
+	// runs even if another node has more free capacity.
+	nodes := []NodeInfo{
+		{Hostname: "big", Runtimes: []string{"process"}, Kind: "linux", CPU: 8, Memory: 8192, FreeCPU: 8, FreeMem: 8192},
+		{Hostname: "small", Runtimes: []string{"process"}, Kind: "linux", CPU: 4, Memory: 4096, FreeCPU: 4, FreeMem: 4096},
+	}
+	redisReq := WorkloadRequest{Spec: serviceSpec("redis", "process", 1, nil), Namespace: "ns"}
+	redisPlacements, err := Schedule(nodes, redisReq)
+	if err != nil {
+		t.Fatalf("redis Schedule: %v", err)
+	}
+	// Redis lands on "big" (most free capacity). Now schedule the
+	// worker with affinity to redis; it should also land on "big".
+	workerReq := WorkloadRequest{Spec: &jobspec.WorkloadSpec{
+		Kind:    "Job",
+		Name:    "worker",
+		Count:   1,
+		Runtime: &jobspec.RuntimeBlock{OneOf: "process"},
+		Affinity: []jobspec.AffinityRule{
+			{Target: "redis", Weight: 1000},
+		},
+	}, Namespace: "ns"}
+	// The affinity is name-based; we need to seed the worker schedule
+	// with the redis placement so affinityScore can see it. Schedule
+	// does not take prior placements, so test affinityScore directly.
+	got := affinityScore(nodes[0], workerReq, redisPlacements)
+	if got <= 0 {
+		t.Errorf("affinityScore(big) = %d, want > 0 (redis colocated)", got)
+	}
+	gotSmall := affinityScore(nodes[1], workerReq, redisPlacements)
+	if gotSmall != 0 {
+		t.Errorf("affinityScore(small) = %d, want 0 (redis not colocated)", gotSmall)
+	}
+}
+
+func TestSchedule_AffinityCELExpression(t *testing.T) {
+	// Affinity with a CEL target: prefer nodes tagged "ssd".
+	nodes := []NodeInfo{
+		{Hostname: "hdd", Runtimes: []string{"process"}, Tags: []string{"hdd"}, Kind: "linux", CPU: 8, Memory: 8192, FreeCPU: 8, FreeMem: 8192},
+		{Hostname: "ssd", Runtimes: []string{"process"}, Tags: []string{"ssd"}, Kind: "linux", CPU: 4, Memory: 4096, FreeCPU: 4, FreeMem: 4096},
+	}
+	req := WorkloadRequest{Spec: &jobspec.WorkloadSpec{
+		Kind:    "Job",
+		Name:    "db",
+		Count:   1,
+		Runtime: &jobspec.RuntimeBlock{OneOf: "process"},
+		Affinity: []jobspec.AffinityRule{
+			{Target: `"ssd" in node.tags`, Weight: 10000},
+		},
+	}, Namespace: "ns"}
+	got, err := Schedule(nodes, req)
+	if err != nil {
+		t.Fatalf("Schedule: %v", err)
+	}
+	if got[0].Node != "ssd" {
+		t.Errorf("Node = %q, want ssd (affinity to ssd tag outweighs capacity)", got[0].Node)
+	}
+}
+
+// ---------------------------------------------------------------------------
+// Error paths
+// ---------------------------------------------------------------------------
+
+func TestSchedule_EmptyNodes(t *testing.T) {
+	req := WorkloadRequest{Spec: jobSpec("x", "process", nil), Namespace: "ns"}
+	if _, err := Schedule(nil, req); err == nil {
+		t.Fatal("Schedule: expected error for empty nodes, got nil")
+	}
+}
+
+func TestSchedule_NilSpec(t *testing.T) {
+	if _, err := Schedule(threeLinuxNodes(), WorkloadRequest{}); err == nil {
+		t.Fatal("Schedule: expected error for nil spec, got nil")
+	}
+}
+
+func TestSchedule_UnknownKind(t *testing.T) {
+	req := WorkloadRequest{Spec: &jobspec.WorkloadSpec{Kind: "Cron", Name: "x", Count: 1}, Namespace: "ns"}
+	if _, err := Schedule(threeLinuxNodes(), req); err == nil {
+		t.Fatal("Schedule: expected error for unknown kind")
+	}
+}
+
+// ---------------------------------------------------------------------------
+// Score unit tests
+// ---------------------------------------------------------------------------
+
+func TestScore_FitsAndDoesNotFit(t *testing.T) {
+	node := NodeInfo{Hostname: "n", Runtimes: []string{"process"}, Kind: "linux", CPU: 4, Memory: 4096, FreeCPU: 4, FreeMem: 4096}
+	req := WorkloadRequest{Spec: jobSpec("j", "process", nil), Namespace: "ns"}
+	score, fits := Score(node, req)
+	if !fits {
+		t.Error("fits = false, want true")
+	}
+	if score <= 0 {
+		t.Errorf("score = %d, want > 0", score)
+	}
+}
+
+func TestScore_RuntimeMismatchDoesNotFit(t *testing.T) {
+	node := NodeInfo{Hostname: "n", Runtimes: []string{"process"}, Kind: "linux", CPU: 4, Memory: 4096, FreeCPU: 4, FreeMem: 4096}
+	req := WorkloadRequest{Spec: jobSpec("j", "wasm", nil), Namespace: "ns"}
+	if _, fits := Score(node, req); fits {
+		t.Error("fits = true for wasm on process-only node, want false")
+	}
+}
+
+func TestScore_ConstraintFailsDoesNotFit(t *testing.T) {
+	node := NodeInfo{Hostname: "n", Runtimes: []string{"process"}, Kind: "linux", CPU: 4, Memory: 4096, FreeCPU: 4, FreeMem: 4096}
+	req := WorkloadRequest{Spec: jobSpec("j", "process", []string{`node.kind == "proxmox"`}), Namespace: "ns"}
+	if _, fits := Score(node, req); fits {
+		t.Error("fits = true for kind==proxmox on linux node, want false")
+	}
+}
+
+// ---------------------------------------------------------------------------
+// allocID / isAllocFor helpers
+// ---------------------------------------------------------------------------
+
+func TestAllocID(t *testing.T) {
+	req := WorkloadRequest{Spec: &jobspec.WorkloadSpec{Name: "web"}, Namespace: "prod"}
+	if got := allocID(req, 2); got != "prod/web-2" {
+		t.Errorf("allocID = %q, want prod/web-2", got)
+	}
+	req.Namespace = ""
+	if got := allocID(req, 0); got != "default/web-0" {
+		t.Errorf("allocID = %q, want default/web-0", got)
+	}
+}
+
+func TestIsAllocFor(t *testing.T) {
+	cases := []struct {
+		allocID  string
+		workload string
+		want     bool
+	}{
+		{"ns/redis-0", "redis", true},
+		{"ns/redis-12", "redis", true},
+		{"ns/worker-0", "redis", false},
+		{"redis-0", "redis", true},
+		{"ns/web-canary-3", "web-canary", true},
+	}
+	for _, c := range cases {
+		if got := isAllocFor(c.allocID, c.workload); got != c.want {
+			t.Errorf("isAllocFor(%q,%q) = %v, want %v", c.allocID, c.workload, got, c.want)
+		}
+	}
+}
+
+// ---------------------------------------------------------------------------
+// normalizeRuntime / hasRuntime
+// ---------------------------------------------------------------------------
+
+func TestHasRuntimeAliases(t *testing.T) {
+	cases := []struct {
+		name string
+		node NodeInfo
+		want bool
+	}{
+		{"wasm on wasmtime node", NodeInfo{Runtimes: []string{"wasmtime"}, Kind: "linux"}, true},
+		{"wasm on process node", NodeInfo{Runtimes: []string{"process"}, Kind: "linux"}, false},
+		{"pve-vm on linux node", NodeInfo{Runtimes: []string{"pve-vm"}, Kind: "linux"}, false},
+		{"pve-vm on proxmox node", NodeInfo{Runtimes: nil, Kind: "proxmox"}, true},
+		{"process on process node", NodeInfo{Runtimes: []string{"process"}, Kind: "linux"}, true},
+		{"empty runtime on any node", NodeInfo{Runtimes: []string{"process"}, Kind: "linux"}, true},
+	}
+	for _, c := range cases {
+		if c.name == "empty runtime on any node" {
+			// hasRuntime is only called when OneOf != "".
+			continue
+		}
+		if got := hasRuntime(c.node, "wasm"); c.name == "wasm on wasmtime node" || c.name == "wasm on process node" {
+			if got != c.want {
+				t.Errorf("%s: hasRuntime(wasm) = %v, want %v", c.name, got, c.want)
+			}
+		}
+	}
+	// Explicit pve-vm and process checks.
+	if !hasRuntime(NodeInfo{Runtimes: nil, Kind: "proxmox"}, "pve-vm") {
+		t.Error("pve-vm on proxmox node should fit")
+	}
+	if hasRuntime(NodeInfo{Runtimes: nil, Kind: "linux"}, "pve-vm") {
+		t.Error("pve-vm on linux node should not fit")
+	}
+	if !hasRuntime(NodeInfo{Runtimes: []string{"process"}, Kind: "linux"}, "process") {
+		t.Error("process on process node should fit")
+	}
+}