feat: NVIDIA Sr DevOps application, drop cohort caps, fix Workday/decision-log bugs

Scout run 2026-09-04 (36 companies, 778 eligible) plus the NVIDIA package it
surfaced, and the tooling and policy fixes found along the way.

Application
- NVIDIA Senior DevOps Engineer, Zurich (JR2024256): submitted 2026-09-05.
  Evidence Fit 64/100 Stretch, hard gate FAIL x2, channel Weak. Sent on an
  explicit user override with the working model unresolved, so hybrid-vs-onsite
  is a live screening topic. 2pp resume + 1pp cover letter, both validators
  PASS, 0 boxes, 0 Tier 1 findings open.

Policy
- Drop the 7 Core / 2 Adjacent / 1 Stretch cohort caps. They had no derivation
  and were overridden whenever they bound, and the cohort was confounded: it
  tested fit class while holding channel constant at Weak across all six
  applications. Keep the recording, drop the rationing.
- Channel replaces the slot budget as the rationed variable: no application
  goes out cold when a named contact is published and reachable.
- Propagated through shared_ops, session_file_template, cohort_tracker
  (counts, never counts-against-a-target) and application_cohort.json.

Tool fixes
- fetch_workday emitted 404 URLs for every Workday company (NVIDIA, Novartis,
  Equinor): externalPath is site-relative and needs the locale + site prefix.
  6/6 short-form 404, 3/3 long-form 200. 773 dead links repaired across 42
  archived reports.
- find_decision_url silently de-linked decisions on two boards, which is why
  ruled-out roles kept resurfacing as undecided. RUAG rotates its UUID on
  repost; BIS migrated its vacancy URL scheme, orphaning 5 decisions including
  a live application. Fixed as a two-tier match, not a slug key -- slugs
  collide (two live RUAG reqs shared one).

Knowledge base
- claims.json: add Bash / shell scripting, the third canonical gap of this
  shape after Ansible and Linux administration.
- bundle_data_platform: BS-4 was framed as a delivered observability stack over
  what claims.json calls a proof of concept -- the defect the 2026-08-21 sweep
  was told to look for.
- CLAUDE.md and AGENTS.md: correct the RUAG Data Lakehouse shortlist claim,
  which the JD does not support and which both files still asserted.
- resume_template: comment-guard the missing babel/ngerman language block.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QFb8tTCig5r1KzkQiNXcYt

resume_builder/reference/application_strategy.md

@@ -27,19 +27,56 @@ Before generating a resume, complete Evidence Fit from `critique_framework.md`.
 
 Do not spend a full resume/cover-letter cycle on a no-go role. Record the decision and move on.
 
-## Controlled Ten-Application Cohort
+## Application Record (no quotas)
 
-The next diagnostic cohort uses:
+**Policy change, 2026-09-05, user decision: the 7 Core / 2 Adjacent / 1 Stretch caps are DROPPED.**
 
-- 7 Core applications.
-- 2 Adjacent applications.
-- 1 Stretch application maximum.
+The caps were arbitrary -- no derivation existed for those numbers -- and they were routed around
+whenever they bound (RUAG AI Engineer went in at Evidence Fit 74 with a hard-gate FAIL; Citadel was
+recorded Adjacent when the strict reading said Stretch). A rule that is overridden whenever it binds
+governs nothing and only adds friction.
 
-Track fit class, hard gates, channel, application date, stage and outcome. Evaluate the cohort only after ten applications or a meaningful screen/interview result; do not infer profile quality from a few high-variance FAANG outcomes.
+The cohort was also a confounded experiment. It was designed to test **fit class**, but the variable
+it silently held constant was **channel**: all six applications were Weak/cold. It therefore could not
+distinguish "Stretch does not convert" from "cold does not convert", and the results point away from
+fit class anyway -- the highest-scoring package in the log (Snowflake, ~86) drew no reply in 80 days,
+while the one interview (SBB) came from a cold Core submit whose resume was *missing* the three ATS
+literals the session had flagged as risky.
+
+**What is kept: the recording, not the rationing.** Every application still records fit class, hard
+gate, channel, date, stage and outcome. That record is what makes patterns visible instead of
+re-derived, and it produced every useful finding to date. `application_cohort.json` continues as a
+durable log; `cohort_tracker.py summary` remains a read-only view. There is no slot to consume and no
+application is blocked for lack of one.
+
+**Fit class is now a preparation signal, not a spend limit.** An Evidence Fit of 64 does not mean
+"do not apply" -- it means know which two questions you will be asked and have honest answers ready.
+Hard gates and NO-GO triggers are unchanged: a required Gap or an unresolved practical constraint is
+still a real finding that must be surfaced before the user decides.
+
+### Real constraints that survive independently of any quota
+
+These are not cohort discipline and must not be dropped with it:
+
+- **Google: 3 applications per rolling 30 days**, and a 90-day wait before reapplying to the *same*
+  role. An external, enforced limit.
+- **Do not stack a second cold application on one organisation while the first is pending** -- a
+  judgment call, made deliberately before (Snowflake Metrics was paused for exactly this).
+- **Each package costs real hours.** Volume is limited by effort, not by policy.
 
 ## Channel Strategy
 
-For every Core application, attempt at least one warm action before or immediately after applying:
+**Channel is the rationed variable, replacing the old slot budget** (2026-09-05). It is the constraint
+with a plausible effect and an untested one: the record stands at **six consecutive cold submits, one
+interview**, with zero Strong or Moderate channels ever attempted.
+
+**Rule: no application goes out cold when a named contact is published and reachable.**
+
+This has been available and unused repeatedly -- Marco Heinzen at RUAG (published direct line), Andri
+Wienandts at SBB (published number, and it still converted), Per Olav Marthinsen at Aker BP. Making
+contact costs one phone call or one short email and is the cheapest experiment left.
+
+For every application, attempt at least one warm action before or immediately after applying:
 
 - credible internal referral;
 - targeted note to a relevant recruiter;
@@ -47,6 +84,7 @@ For every Core application, attempt at least one warm action before or immediate
 - former colleague/customer/vendor connection who can establish work quality.
 
 Record the channel as Strong, Moderate or Weak. Tailoring alone does not upgrade a cold channel.
+When no contact is published anywhere, a cold submit is fine -- record it as Weak and move on.
 
 ## Resume Strategy
 
@@ -63,10 +101,10 @@ Apply the decision gate in `cl_reference.md`. A letter is not a default package
 
 ## Feedback Loop
 
-After the ten-application cohort, compare:
+Review the record whenever a meaningful screen or interview result lands, and compare:
 
+- Warm vs cold progression -- **the primary question now, and still unanswered.**
 - Core vs Adjacent/Stretch progression.
-- Warm vs cold progression.
 - Recruiter rejection vs hiring-manager rejection.
 - Tool/domain patterns among successful and unsuccessful roles.
 - Whether resume changes affected outcomes.