feat: add multi-phase workflow enforcement rules for skill generation (#236)

* feat: add multi-phase workflow enforcement rules for skill generation

The agent-instructions format and author-agent-instructions template had
no guidance for multi-phase interactive workflows with explicit gating.
Skills generated from complex pipelines (e.g., a 7-pass PCB design
workflow) would describe the pipeline architecturally but not
behaviorally constrain the agent, causing it to skip intermediate steps
and jump directly to final output generation.

Add 'Multi-phase workflow rules' to the agent-instructions format
covering four required enforcement mechanisms:
- Critical constraints block (prohibited output types, self-correction)
- Phase gate directives (explicit stop at each phase boundary)
- Pass tracking block (observable state announcement)
- Anti-shortcut warning (failure mode rationale)

Update the author-agent-instructions template to:
- Detect multi-phase source templates and apply the new rules
- Preserve phase gating logic instead of condensing it away
- Add Critical Rule blocks at each phase boundary

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: Alan Jowett <alan.jowett@microsoft.com>

* feat: add IR consumption rule for generated scripts in multi-phase workflows

Scripts produced by post-binding tool steps (e.g., schematic renderers,
PCB layout generators) were hardcoding design data as Python dicts
instead of reading from the IR files that earlier passes produced. This
defeats the purpose of the IR architecture — the IRs exist but are
ignored.

Add rule 5 (IR consumption rule) to the multi-phase workflow rules:
- Generated code MUST read component lists, net names, values, and
  library mappings from IR files at runtime
- Only format-specific syntax constants are permitted as hardcoded values
- Self-check: if deleting the IR files would not change the script's
  output, the script is hardcoding data — rewrite it

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: Alan Jowett <alan.jowett@microsoft.com>

* feat: propagate IR consumption rule to author-agent-instructions template

The template's Step 2 (skill planning) and Step 3 (condense) referenced
the multi-phase workflow rules but didn't mention the IR consumption
rule added in the previous commit. Both locations now include the
directive: generated code must read from artifacts at runtime, never
hardcode data, with the self-check.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: Alan Jowett <alan.jowett@microsoft.com>

* feat: add fail-stop backstop to IR consumption rule

If generated code cannot read or parse a required artifact (file
missing, malformed YAML, missing fields), it MUST raise an error and
halt. It MUST NOT fall back to generating data from memory, prior
context, or hardcoded defaults. A missing artifact means an earlier
pass needs re-execution, not a workaround.

This closes the last loophole: without this rule, an agent hitting a
friction point reading an IR file could silently fall back to inlining
the data from conversation context, which is how hardcoding starts.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: Alan Jowett <alan.jowett@microsoft.com>

* feat: add field-level mapping and gate-driven iteration rules

Two additions to the IR consumption rule:

1. Field-level mapping: Tool steps that produce scripts must enumerate
   which specific artifact fields the script reads (e.g., connector
   positions from connector_placement[].position.x_mm), not just
   generic references like 'read from IR-3'. Generic references produce
   scripts that ignore detailed fields.

2. Gate-driven iteration: When a tool step has a validation gate (e.g.,
   DRC zero violations) and the gate fails, the agent must iterate on
   the script to fix violations — not ask the user whether to proceed.
   The agent only presents output for user approval after gate criteria
   are satisfied. Asking the user to accept a failing gate is itself a
   gate violation.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: Alan Jowett <alan.jowett@microsoft.com>

* fix: address PR review comments — consistency and formatting

- Convert 'Multi-phase workflow rules' from bold text to ### heading
  for stable anchor linking
- Align multi-phase trigger criteria between format and template:
  '3+ sequential phases' or 'interactive with explicit gates/passes'
- Fix fenced code block indentation in list items 2 and 3 for correct
  GitHub Markdown rendering (4-space indent + blank line separation)
- Align Critical Constraints placement wording: 'immediately after the
  opening task description, before architecture/methodology content'

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: Alan Jowett <alan.jowett@microsoft.com>

* fix: align Critical Constraints heading with format spec

Use the full heading text 'Critical Constraints — Read Before Doing
Anything' as recommended by the format spec to maximize salience and
reduce variability in generated skills.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: Alan Jowett <alan.jowett@microsoft.com>

* fix: address round 4 review comments

- Demote Multi-phase workflow rules from ### to #### so it nests
  under section 4 (CLI Skill) and preserves the document outline
- Expand the IR consumption rule checklist in the template's Step 2
  to include all sub-requirements: fail-stop, field-level mapping,
  self-check, and gate-driven iteration

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: Alan Jowett <alan.jowett@microsoft.com>

---------

Signed-off-by: Alan Jowett <alan.jowett@microsoft.com>
Co-authored-by: Alan Jowett <alan.jowett@microsoft.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>